IMP logo
IMP Manual  for IMP version 2.11.0
ChangeLog.md
1 ChangeLog {#changelog}
2 =========
3 
4 # 2.11.0 - 2019-06-25 # {#changelog_2_11_0}
5 - A number of new tutorials are now available covering a variety of topics
6  from modeling complexes with IMP::pmi to writing new C++ code. See
7  the [tutorial index](https://integrativemodeling.org/tutorials/) for
8  more details.
9 - The build system has been updated to use new CMake functionality
10  (if available) to detect Python. If you have CMake 3.12 or later, this should
11  do a much better job of handling multiple Python installations (e.g. Macs
12  with Homebrew). It will default to using Python 3. To force use of Python 2,
13  pass `-DUSE_PYTHON2=on` to your CMake invocation.
14 - CMake files are now provided to make it straightforward to use IMP as
15  a library in other programs. See
16  [the tutorial](https://integrativemodeling.org/tutorials/using_cpp/develop/)
17  for more details.
18 - CMake files are now provided to build new IMP modules outside of the IMP
19  build tree (i.e. a module can be built against an IMP binary).
20 - The IMP::OptionParser class has been removed. Use IMP::ArgumentParser instead.
21 - The IMP::test::TempDir and IMP::test::RunInTempDir utility classes have been
22  removed. Use the IMP::test::temporary_directory and
23  IMP::temporary_working_directory context managers instead.
24 
25 # 2.10.1 - 2019-02-26 # {#changelog_2_10_1}
26 - Add support for OpenCV 4.
27 - Fix IMP::isd `create_gmm.py` script to handle command line options correctly.
28 - Command line tools in the Mac and Ubuntu packages should now use system
29  Python (/usr/bin/python), not the first Python (e.g. Anaconda Python)
30  found in PATH (which might not be compatible with IMP's Python libraries).
31 
32 # 2.10.0 - 2019-01-08 # {#changelog_2_10_0}
33 - Applications using IMP::pmi should now use the IMP::pmi::topology::System
34  class to set up the system (sometimes referred to as "PMI2"), rather than
35  the older IMP::pmi::representation::Representation class ("PMI1").
36  This new class is faster and produces standard IMP Hierarchy objects. The
37  old PMI1-style classes are deprecated in this release and will be removed
38  in the next one. For support of published studies using "PMI1", a legacy
39  IMP::pmi1 module is provided (to use it, simply replace all imports of
40  "IMP.pmi" with "IMP.pmi1"). This module will receive only minor bug fixes,
41  and no new features.
42 - All IMP Python command line tools now use
43  [argparse](https://docs.python.org/2/library/argparse.html) rather than
44  [optparse](https://docs.python.org/2/library/optparse.html) for handling
45  options, as optparse is deprecated in modern Python versions. The
46  IMP::OptionParser class, which extends optparse to add standard IMP options,
47  is now deprecated. Use the similar IMP::ArgumentParser class instead.
48 - The bundled copy of [python-ihm](https://github.com/ihmwg/python-ihm) is
49  updated to the latest release. This supports faster (C-accelerated) reading
50  of mmCIF files, read and write support for BinaryCIF files, and support
51  for output of PMI2-style models. The new IMP::atom::read_mmcif function
52  allows reading a PDB file in mmCIF format, similarly to the existing
53  IMP::atom::read_pdb function.
54 - A new application of IMP is now available:
55  - [Architecture of Pol II(G) and molecular mechanism of transcription regulation by Gdown1](https://salilab.org/pol_ii_g/)
56 - The deprecated IMP::isd::FNormal::density() method has been removed.
57  Use IMP::isd::FNormal::get_density() instead.
58 - Handling of quaternions (for example in the calculation of rotational
59  derivatives) is now faster and more numerically stable.
60 - IMP::mpi now includes a workaround for an OpenMPI failure when used
61  with Python, so previous hacks (building OpenMPI with the
62  `--disable-dlopen` flag, or using `sys.setdlopenflags` in Python)
63  should no longer be necessary.
64 
65 # 2.9.0 - 2018-07-02 # {#changelog_2_9_0}
66 - [Eigen](https://eigen.tuxfamily.org/) is now required to build IMP.
67 - The Windows .exe IMP installer no longer supports Python 2.6. Python 2.7,
68  3.4, 3.5, and 3.6 are supported by this installer.
69 - A binary package is now provided for Ubuntu 18.04 LTS, Bionic Beaver.
70 - The IMP::atom::setup_as_rigid_body() function has been removed.
71  Use IMP::atom::create_rigid_body() instead.
72 - Constructors for IMP::container classes, IMP::core::BallMover,
73  IMP::core::RigidBodyMover, and IMP::core::IncrementalScoringFunction taking
74  a RigidBody or Particles have been removed. Use constructors taking
75  ParticleIndexes instead.
76 - The IMP::core::ConnectivityRestraint::get_connected_pairs() method has been
77  removed. Use get_connected_index_pairs() instead.
78 - The stats() profiling method has been removed from
79  IMP::isd::GaussianProcessInterpolationRestraint and
80  IMP::isd::MultivariateFNormalSufficient. Use a third party package (such as
81  gperftools) for profiling instead.
82 - Scores in IMP.saxs (and FoXS) now use the chi-square value rather than chi.
83 - New applications of IMP are now available:
84  - [Modeling of the entire yeast nuclear pore complex](https://salilab.org/npc2018), as well as
85  [modeling of the flexible FG regions](https://salilab.org/npc_fg_2018) and
86  modeling of the individual [Nup133](https://salilab.org/nup133),
87  [Nup82](https://salilab.org/nup82), and
88  [Pom152](https://salilab.org/pom152) subcomplexes.
89  - [Modeling of the Ecm29-proteasome complex](https://salilab.org/ecm29)
90 - The new modules IMP::npc and IMP::npctransport provide specialized
91  restraints and other functionality for modeling static nuclear pore complex
92  structures and transport through these structures, respectively, although
93  may also be useful for other modeling applications.
94 - IMP's mmCIF support for deposition of models in
95  [PDB-Dev](https://pdb-dev.wwpdb.org/), in the IMP::pmi and IMP::mmcif modules,
96  now utilizes the [python-ihm](https://github.com/ihmwg/python-ihm) library,
97  a copy of which is included with IMP.
98 - RMF files now store additional metadata, such as the primary sequence of
99  each chain, software used, information on each restraint (such as fit values),
100  and provenance (classes derived from IMP::core::Provenance) that describes
101  transformations of the system, such as initial input from a PDB file,
102  sampling, clustering, and filtering.
103 - The new experimental IMP::mmcif module allows generating basic mmCIF files
104  directly from an IMP::Model and/or RMF files.
105 
106 # 2.8.0 - 2017-08-16 # {#changelog_2_8_0}
107 - New applications of IMP are now available:
108  - [Modeling of the yeast Spindle Pole Body core](https://salilab.org/spb/)
109  - [Modeling of the SEA complex, a modulator of the TORC1 pathway](https://salilab.org/sea/)
110 - This will be the last release that includes packages for Ubuntu 12.04
111  (Precise Pangolin) since it has reached end of life. We will continue to
112  provide packages for newer Ubuntu LTS releases.
113 - The Anaconda and Windows .exe packages are now MPI-enabled. On Windows, in
114  order to use MPI, install [MS-MPI](https://msdn.microsoft.com/en-us/library/bb524831.aspx).
115 - IMP::em2d::PCAFitRestraint now uses binary PGM files as output, rather than
116  the less well supported text PGM format. (It can still read both text and
117  binary PGM files as input.)
118 - IMP::em::FitRestraint::cast() is deprecated. Use `get_from()` instead.
119 - The deprecated profiling (timing statistics) support has been removed.
120  A number of third party packages (such as gperftools) are available
121  for profiling and should be used instead.
122 - The new module IMP::spb provides specialized restraints, samplers, and
123  other objects to support modeling of the Spindle Pole Body (SPB). While
124  designed specifically for this system, they may be useful for modeling
125  similar systems.
126 
127 # 2.7.0 - 2017-04-13 # {#changelog_2_7_0}
128 - New applications of IMP are now available:
129  - [Modeling of the yeast Mediator complex](https://salilab.org/mediator/)
130  - [Modeling of the yeast and human TFIIH complex](https://salilab.org/tfiih/)
131  - [Modeling of complexes between the 26S proteasome and proteasome interacting proteins](https://salilab.org/26S-PIPs)
132  - [Human Complement modeling](https://salilab.org/Complement)
133  - [Modeling of a region of the fly genome](https://integrativemodeling.org/systems/?sys=22) using IMP and [TADbit](http://sgt.cnag.cat/3dg/tadbit/)
134 - Memory and CPU performance improvements across the code, particularly
135  in IMP::pmi.
136 - Preview of mmCIF support for deposition of integrative models in
137  [PDB-Dev](https://pdb-dev.wwpdb.org/). See the IMP::pmi::mmcif
138  module for more details.
139 - This will be the last release that includes packages for RedHat Enterprise
140  Linux 5 (and variants, such as CentOS 5), since RHEL 5 has reached end
141  of life.
142 - The IMP::em::SurfaceShellDensityMap::binaries() method has been replaced
143  with the (correctly spelled) IMP::em::SurfaceShellDensityMap::binarize()
144  method.
145 - IMP::multifit::WeightedExcludedVolumeRestraint no longer expects that the
146  IMP::core::RigidBody particles it acts on are also IMP::atom::Hierarchy
147  objects. Use an IMP::multifit::RigidLeavesRefiner in place of
148  IMP::core::LeavesRefiner to get the old behavior.
149 - IMP::modeller::MultipleBinormalRestraint has moved to the IMP::core
150  module.
151 - The IMP::core::setup_as_rigid_body() method has been removed.
152  Use IMP::core::create_rigid_body() instead.
153 - IMP::test:TempDir and IMP::test::RunInTempDir are deprecated in favor
154  of similar functions that work as Python context managers
155  (IMP::test::temporary_directory() and IMP::test::temporary_working_directory()
156  respectively).
157 
158 # 2.6.2 - 2016-05-25 # {#changelog_2_6_2}
159 - Add support for SWIG 3.0.8.
160 - Add support for Ubuntu 16.04, Xenial Xerus.
161 - Minor documentation fixes.
162 - Bugfix: fix detection of termini in IMP::atom::Selection.
163 
164 # 2.6.1 - 2016-04-14 # {#changelog_2_6_1}
165 - The 'cluster' binary is now called 'rmsd_cluster' to avoid conflicts with
166  other packages (e.g. graphviz also provides a 'cluster' binary).
167 - Minor bug fixes to PMI.
168 - Minor documentation fixes.
169 
170 # 2.6.0 - 2016-03-28 # {#changelog_2_6_0}
171 - This release includes most of the functionality of PMI2. This is an improved
172  version of PMI that is faster than PMI1, more modular, and easier to use.
173  See IMP::pmi for more information.
174 - A new application of IMP is now available:
175  [Modeling of exosome complexes from cross-link MS data](http://salilab.org/exosome).
176 - The IMP::base and IMP::kernel namespaces have been removed (use the top-level
177  IMP namespace instead).
178 - The IMP::display::ChimeraWriter class has been removed (use IMP::rmf instead).
179 - IMP::Model no longer has an implicit scoring function. This means that
180  Model::evaluate() and Model::add_restraint() have been removed (use an
181  IMP::ScoringFunction instead). IMP::Optimizer and IMP::Sampler now also
182  require an IMP::ScoringFunction to be explicitly set before they are used.
183 - IMP::atom::Hierarchy::get_as_xxx() methods are deprecated; use regular
184  decorators instead. For example, instead of h.get_as_residue() use
185  Residue(h). If you want to check for validity first,
186  use Residue::get_is_setup(h).
187 - IMP::atom::Selection can now select by element.
188 - IMP::modeller::load_restraints_file() and IMP::modeller::read_pdb() have
189  been removed. Use the IMP::modeller::ModelLoader class instead.
190 - IMP::atom::PDBSelector subclasses are more flexible, with a new
191  IMP::atom::XorPDBSelector class, and can be combined using standard
192  Python operators, in similar fashion to sets (&, |, ^, -, ~ operators).
193 - The following modules that use FFTW (which is licensed under the GPL)
194  are themselves now also licensed under the GPL: IMP::EMageFit,
195  IMP::cnmultifit, IMP::em2d, IMP::integrative_docking, and IMP::multifit.
196 - Bugfix: sampled EM density maps should now handle resolution correctly; maps
197  generated by the fixed code agree well with those generated by EMAN. The old
198  radius-dependent kernel parameter code is no longer used.
199 
200 # 2.5.0 - 2015-06-30 # {#changelog_2_5_0}
201 - Documentation has been revised and updated; the main manual now gives a
202  gentler introduction to IMP, with more technical details on the code
203  provided in a separate reference guide.
204 - The IMP::base and IMP::kernel namespaces are now deprecated. Functionality
205  in both base and kernel now lives in the top-level "IMP" namespace.
206 - IMP::atom::Selection::get_selected_particle_indexes() has a new
207  'with_representation' parameter. If turned on (the default), Selection
208  returns all representational (XYZ) particles that match the request -
209  for example, selecting a residue will typically return all of its atoms.
210  If turned off, Selection returns the highest matching particle instead -
211  so, in this case, the residue particle itself.
212 - All IMP::RAII objects (e.g. IMP::SetLogState) now support the context
213  manager protocol in Python (e.g. "with IMP.SetLogState() as s:")
214 - Several long-deprecated functions have been removed. These include:
215  - Several Restraint constructors that take Particles. Use equivalent
216  constructors that take an IMP::Model pointer and ParticleIndexes instead.
217  - IMP::core::MonteCarlo::get_number_of_forward_steps() (use
218  get_number_of_accepted_steps() instead).
219  - IMP::Model::remove_restraint(), IMP::Model::get_restraints(),
220  IMP::Model::get_number_of_restraints(), IMP::Model::get_restraints().
221  Store your restraints in an IMP::RestraintSet or use an
222  IMP::core::RestraintsScoringFunction. Related methods such as
223  IMP::Model::evaluate() will be removed in the next IMP release.
224  - IMP::Model::set_maximum_score() and related methods. Set the maximum
225  scores on the IMP::Restraint or IMP::RestraintSet directly, using
226  IMP::Restraint::set_maximum_score().
227  - IMP::Restraint::get_input_containers() and
228  IMP::Restraint::get_input_particles(). Use IMP::Restraint::get_inputs()
229  instead (similarly for get_outputs()).
230  - IMP::rmf::save_frame() no longer takes a frame index argument (it was
231  previously ignored anyway).
232 - Several functions have been deprecated in this release (and will be removed
233  in the next). These include several Restraint constructors that take Particle
234  or Particles objects (use constructors that take an IMP::Model and
235  ParticleIndex or ParticleIndexes instead); and add_particle*(),
236  set_particle*() and clear_particle*() in several Container objects (use add(),
237  set() and clear() with particle indexes instead).
238 - On platforms that support it, OpenMP is turned on by default. To speed up
239  some parts of IMP, call IMP::base::set_number_of_threads(). By default,
240  only a single thread is used.
241 - The RPM packages now include an -mpich subpackage that can be installed
242  to enable parallel computation, using the IMP.mpi module in concert
243  with the mpich library.
244 - We now provide packages for both of the currently-supported Ubuntu LTS
245  releases: Ubuntu 12.04 (Precise Pangolin) and Ubuntu 14.04 (Trusty Tahr).
246 - The default log level has changed from NONE to WARNING.
247 - A new command line tool, `multi_foxs`, allows for enumeration and scoring
248  of multi-state models against SAXS profiles.
249 - A new command line tool, `ligand_score_multiple`, acts similarly to the
250  existing `ligand_score` tool but scores each ligand pose generated by
251  docking software (such as PatchDock).
252 - New applications of IMP are now available:
253  - [Multi-state modeling of the histidine kinase, PhoQ](http://salilab.org/phoq).
254  - [Determination of the molecular architecture of the 40Sâ˘eIF1•eIF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3). - IMP::domino::DiscreteSampler::set_restraints() now allows the set of restraints used in DOMINO filtering to be controlled (previously, all restraints in the IMP::Model were used). - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the MODELLER score, not the implicit (and deprecated) score of the entire IMP::Model. - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently treated (they are supposed to be the centers of the voxels, but in some cases they were the lower left edge). The fix may result in maps and their bounding boxes appearing shifted by half the voxel size. # 2.4.0 - 2015-02-27 # {#changelog_2_4_0} - Python 3 is now supported. The Windows package now works out of the box with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu and Fedora. (Apple does not yet provide Python 3 so the Mac package is not built with Python 3 support, but if you install Python 3 you can compile IMP from source to work with it.) - A new command line tool, `rrt_sample`, allows for running rapidly exploring random tree (RRT) based sampling, e.g. to sample conformations of multiple domains connected by flexible linkers. - Command line tools are now placed in the `bin` directory of modules (the old `bin` directory, used for utility programs that are not installed, is now called `utility`). The `applications` directory is removed. - Several Python modules that were only used by EMageFit have been moved from IMP::em2d to IMP::EMageFit. - A new tutorial that uses the IMP::pmi module is now available (modeling of the RNA Polymerase II stalk). # 2.3.1 - 2014-12-01 # {#changelog_2_3_1} - Minor bug fixes. - Update to IMP::pmi to support the Nup84 and rnapolii biological systems. # 2.3.0 - 2014-10-24 # {#changelog_2_3_0} - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental. - Binary installers for 64-bit Windows are now available. - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications. - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used. - The integrative docking (idock) application now employs SOAP to improve the scoring. - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments. - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python. # 2.2.1 - 2014-08-11 # {#changelog_2_2_1} - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/) # 2.2.0 - 2014-04-09 # {#changelog_2_2_0} - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that. - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back. - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry(). - The IMP::base::Flag class was added to simplify adding flags in C++. - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers. - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones. - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls. - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application. - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead. - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added. - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available. - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples. - The container implementations has been cleaned up and simplified a bit. This should not change most usage. - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`). - [Eigen](#Eigen) is now included with IMP. - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead. - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose. - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer. - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h. - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string. - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/. # 2.1.1 - 2013-11-04 # {#changelog_2_1_1} - Added support for Mac OS 10.9 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0} - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time. - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it. - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member. - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before). - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc. - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput. - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation. - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before. - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs. - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages. - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`). - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on. - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested. - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up. - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized. # For IMP 2.0 and earlier # {#changelog_2_0} ## Spring 2013 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned. - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module. - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory. - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS. ## Winter 2012-2013 - Scons support was removed. Use cmake to build IMP, it is way better. - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long. - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody. - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols. - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo. - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems. - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake). - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit. - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file. - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice. - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript. - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case. - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed. - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info. - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far. - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin. - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code. - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page. - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python. - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated. - IMP::base::FailureHandler has been removed as it wasn't used. - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer. - Modules now must be in a folder called `modules`, which must have a `description.py` file. - The build system should now be a little better about cleaning up after itself. - How headers/doc files/data/examples - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name. - Nice version numbers will now be generated if you use git. You need to rerun tools/git/developer_tools/init_git if you have an existing clone. ## Fall 2012 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed. - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro. - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively. - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason. - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know. - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting. - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles. - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container. - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this. - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes. ## Summer 2012 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro. - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API. - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons. - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors. - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling. - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names. ## Winter 2011/2012 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy. - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()). - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes. - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores. - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input). - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use. - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented. - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed. - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects. - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct. - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole. - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed. - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer). - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions. - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified. - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated. - 24/1: \c tools/make-source has been added that adds a source/header pair to a module - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD. - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use. - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations. - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well. - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided) - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values. ## Fall 2011 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector). - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector. - IMP::statistics::HistogramD now supports display with \c matplotlib. - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply. - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn. - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom. - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel. - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d() - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support. - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility. ## Summer 2011 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy. - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone. - IMP::core::ExcludedVolumeRestraint now supports filters - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before. - you can now pick which Python to use with the \c python argument to scons. - the units for BD diffusion coefficients changed to be consistent with other IMP units. - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb. - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform. - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types. - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably. - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated. - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code. - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule. - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step. - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it. - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors. - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants. - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster. - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values. - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference(). - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially faster when used with IMP::Model::evaluate_if_good() - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value. - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder. - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent. - the RMF helper binaries have been moved to their own application so that they get installed. - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written. ## Spring 2011 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster. - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute domino on a cluster as well as debugging the sampling process. - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf. - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz. - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints. - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support. - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them. - The Python graph interface now has a remove_vertex method. - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState. - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier. - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes. ## Winter 2011 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step() - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing. - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful). - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD. - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score. - each IMP::Restraint now has its own weight. This is lightly tested. - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward. - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested). - the name of the clustering methods have been changed to create_ from get_ since they create objects. - the names of the recently added metric based clustering support has been changed to Metric rather than Distance. - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript. - the fuction IMP::display::create_restraint_geometry() was added to do what its name says. - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches. - IMP::atom::DopePairScore has been added to IMP. - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified interface to the bonded part of the CHARMM forcefield. - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons. - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide. - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and residue naming to PDB-style, so CHARMM topology files can be used unmodified. - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes. - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information. - IMP::atom::CHARMMTopology::add_missing_atoms() and IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms read from a PDB file match the CHARMM topology. - building of deprecated functionality is now false by default - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP. - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath). - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead. - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles. - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP. - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is. - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call. - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists. - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info. - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value. - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards. - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this means is that they need to be created with \c new in C++ code. It also means that one is no longer prohibited from storing them (not that there is much use in doing that). - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things. - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better. - the transform based methods on reference frame which have been deprecated for a while were removed - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source. - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version ## Fall 2010 - IMP::display::PlaneGeometry was added to display planes - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to benefit from better checking. - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing. - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer() - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState. - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD. - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory. - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that. - Logging has been improved to better report the name of methods producing the log messages - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output. - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric. - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest. - applications and biological systems now have their own documentation pages - IMP::atom::Hierarchy requires radii on its leaves - IMP::atom::read_pdb add radii to the atoms - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead. - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too. - a new example was added in the kernel to show IMP used on a biological system - IMP::display::PymolWriter now supports adding labels to things - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints. - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions. - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members. - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring. - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls ## Summer 2010 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work. - Javi added his 2D EM scoring functionality as IMP::em2D - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually. - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer). - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc. - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data). - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process - a better explanation of how to use reference counted pointers has been added to IMP::Pointer. - all simple, planar polygons are now supported for display when CGAL is present ## Week of April 1, 2010 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are - types are not checked until the lists are passed to C++ - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python) - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used. - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work. - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying. ## Week of March 13, 2010 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere. - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors). ## Week of March 6, 2010 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite. ## Week of Feb 27, 2010 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis(). - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things - renamed IMP::core::root() to IMP::core::get_root() ## Week of Feb 21, 2010 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means. - various more interpolation and endian methods had their names fixed (by adding \c get_) - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid() - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d() - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions. - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>). - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points. Some good sed strings are - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_in_box(\#get_random_vector_in(\#g" - "s\#random_vector_on_box(\#get_random_vector_on(\#g" - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g" - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g" - "s\#grid_cover\#get_grid_surface_cover\#g" - "s\#uniform_coverr\#get_uniform_surface_cover\#g" . - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD. - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments. - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis() - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point) - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base. - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance() The following global replacements are probably safe (in sed expressions) - "s\#core.distance(\#core.get_distance\#g" - "s\#algebra.distance(\#algebra.get_distance\#g" - "s\#\.transform(\#.get_transformed(\#g" Only in C++ - "s\#\.rotate(\#.get_rotated(\#g" Only in C++ - "s\#rotation_from\#get_rotation_from\#g" - "s\#rotation_in\#get_rotation_in\#g" - "s\#identity_rotation(\#get_identity_rotation_3d(" - "s\#::distance(\#::get_distance(\#g" . - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry. - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future. - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair() ## Week of Feb 14, 2010 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support. - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged. - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use. - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector. - IMP::example renamed (from examples) - The IMP_USAGE_CHECK() macro lost the (unused) third argument. ## Week of February 3, 2010 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model() - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically. - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object. - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show(). - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin. - IMP.atom.create_protein() now returns the created protein - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone(). - IMP.helper.clone() has been renamed to IMP.atom.create_clone() - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner. - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added. - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages. - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement. - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used. - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume(). ## Week of January 22, 2010 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(), IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds. - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer. - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles. - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint. - Add IMP::core::MinumumPairScore (and kin). ## Week of January 1, 2010 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed. - IMP::UsageException and IMP::InternalException are now reserved for their respective checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API exceptions (which are thrown in any build mode and should be documented and tested) and helper exceptions which are only thrown when checks are turned on (and hence should not be documented) clearer. - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves. This means that incremental evaluation of restraints in restraint sets now works and that restraints in restraint sets can be individually evaluated. Existing working usage should not have changed (so \c sed should manage updates just fine), but the internal implementation is completely new, so pay attention. - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are convertible to C++ streams and implicitly constructable from both streams and file paths. As a result, functions that take them will work when called with a string for a file path, a Python file or a C++ stream. - The IMP.examples module now documents how to use the needed aspects of scons and SWIG. - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested. - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there. - boost program options is now a checked for dependency. - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them. - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented. - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file). - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas. - a new base class IMP::Constraint has been added for constraints. Various other classes have been renamed to be consistent. ## Week of December 11 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference. - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed. - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation. - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles). ## Week of December 4 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems. - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data. - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more). ## Week of November 19 - imp.xcodeproj was moved into tools and updated accordingly - the bin directory finally went away - IMP::display::ChimeraWriter now supports the latest version of Chimera - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO. - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method. - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors) ## Week of November 12 - All exception types are now mapped to equivalent classes in Python. The class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception in both C++ and Python) so all IMP exceptions can be caught in Python using "except IMP.Exception". For convenience, the IMP::IOException, IMP::ValueException and IMP::IndexException classes in Python also derive from the similarly-named standard Python classes (IOError, ValueError and IndexError respectively). ## week of October 29 Update - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested). - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box(). - Added IMP::core::VolumeRestraint. - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information. - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy(). - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite. - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs. - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such. - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were. - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added. - All containers now keep track of what they looked like on the previous time step. - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier. - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications. - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints. ## Week of October 15 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement. ## Week of October 8 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems. - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this. - IMP_check() and IMP_assert() have been replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK() respectively. Hopefully this makes it clearer when to use which. IMP::set_check_level() has been revised accordingly (to use IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and IMP_not_implemented() have been renamed to IMP_FAILURE() and IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update external code, - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL - replace IMP.CHEAP with IMP.USAGE - replace IMP_assert and IMP_check as above. ## Week of October 1 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used. - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them. - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly. - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too. ## Week of September 27 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs. - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies. - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking. - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins. ## Week of September 20 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier. - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof. - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always. - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version. ## Week of August 30 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles in a box. - str should now work better in Python when applied to IMP::Object objects and a few others. Objects should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show functions directory to make it so that they work in Python. It is now a bug to \c print or \c str to not return something useful in Python for any IMP object. ## Week of August 23 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur. - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names. - Python documentation is now parsed by doxygen ## Week of August 15 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from. ## Week of August 8 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue. - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements). - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that. - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation. - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency. - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate. - The contents of bin are moving to tools since the two directories both contained assorted development tools (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon. - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored). Anyone who implemented a decorator will have to delete the second argument of their show function. ## Week of August 1 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(), IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(), IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply(). - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle() and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which updates the names in all code in subdirectories of the current directory. Please backup things first. - a function, IMP::Restraint::safe_evaluate() has been added to be used when requesting the value of restraints outside of the IMP::Model::evaluate(). ## Week of July 18 - The atom::BondPairContainer has been renamed to atom::BondPairFilter since it wasn't a real container. A real atom::BondPairContainer has been added. - Out of source builds mostly works. To do an out of source build, make a link to \c scons_tools and \c SConstruct from the source directory in the build directory and then add the line \command{repository="../path_to_src_dir"} in the \c config.py in the build directory. Then running \c scons in the build directory will put generated files there. This allows you to make both a \c release and \c fast build off of the same set of sources. Proper docs will be added soon. - The SConscript setup for a module has been greatly simplified so it should be easy enough for users to maintain their own module and updates to the build system shouldn't involve changes to user SConscripts any more. This does mean that any non-svn modules need to be updated. Look at IMP.core or another module and copy what is there. Again, proper docs will be added soon. ## Week of July 11 - Per Javi's suggestion, the conventions have been split into "Coding conventions" and "API conventions" and linked from the main page. - benchmarks got moved to their own module, IMP::benchmark and will now be tracked over time ## Week of July 4 - FilteredListSingletonContainer and FilteredPairSingletonContainer went away as there was too much hard to factor code shared with IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The IMP::core::ClosePairsScoreState is now handled by a list of filters on the class itself. That is, do IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a filter. - Hooks have been added to allow actions to be taken on an assertion or check failure. See IMP::HandleFailure and its descendants for more details. - IMP::display::CGOWriter has been much improved and now allows provides lots of display-time control and easy building of animations. - Generic optimizer states have been added (IMP::core::SingletonsOptimizerState and IMP::core::PairsOptimizerState). ## Week of June 15 - box distributions (IMP::algebra::random_vector_on_box(), IMP::algebra::random_vector_in_box()) take bounding boxes - IMP::em::DensityMap are now IMP::Object objects. Some current allocations on the stack may need to be changed (since ref counted objects can't be stack allocated). - IMP::em::DensityMap::Read() is now deprecated (as is IMP::em::DensityMap::Write) as they violated a slew of conventions. Use IMP::em::read_map() and IMP::em::write_map() instead. ## Week of June 8 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds(); - Hierarchys to IMP::atom::Hierarchies - IMP::Particles has been changed to a new type. It no longer plays quite so well with Python lists (you can't create one from an arbitrary Python list). - The containers for multiple decorators have been revised to make them more useful. See IMP::Decorators. You can now implicitly convert, for %example, a IMP::core::XYZs to an IMP::Particles (as well as use IMP::Decorators::get_particles()). All methods that take a collection of particles with a certain type now instead take a IMP::Decorators. - A method that has expectations about the particles it takes should take decorators rather than particles. That is, anything which expects IMP::core::XYZ particles, should take an IMP::core::XYZs instead of IMP::Particles. - IMP_LIST() now takes an extra argument (the name of the plural container) - In C++ IMP::Decorator objects are implicitly convertible to IMP::Particle* objects. Unfortunately, making this work in Python is problematic. - More methods take and return decorators rather than IMP::Particles. This means it is clearer what their expectations are and what they produce. ## Week of May 15 - The IMP::helper module was added and various functionality moved there. See its documentation page for a list of what is there. - IMP::algebra::SphericalCoords was renamed to IMP::algebra::SphericalVector3D to be consistent with other algebra classes. - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent with the other similar classes. - We are trying to be a bit more specific about what is expected from a IMP::atom::Hierarchy. Towards that end, there are two methods for extracting a representation from a hierarchy, IMP::atom::get_detailed_representation(), and IMP::atom::get_simplified_representation() - A method has been added to produce a simplified copy of an IMP::atom::Hierarchy. The method is still a bit crude and could benefit from other people inspecting the output and saying how it should be improved. - There is now an IMP::algebra::BoundingBoxD (and IMP::algebra::BoundingBox3D) to track bounding boxes. There are methods like IMP::algebra::bound_box(Sphere3D) which gets the bounding box of a sphere. - Each IMP::Object has its own log level. If IMP::Object::set_log_level() is called with a value other than IMP::DEFAULT, the value is used for code within and called by the object. To support this in your objects you need to add IMP_OBJECT_LOG to the start of each major method. It has been added to some of the core objects. ## Week of May 8 - IMP::atom::MolecularHierarchy has been renamed to IMP::atom::Hierarchy. Use \command{sed -i.old "s\#MolecularHierarchy\#Hierarchy\#g" my_files} - the long-deprecated Name(Decorator) was removed - The decorator suffix has been removed from decorator names. - Unit conventions have been established and documented for distance, energy, force and angle. - there is a function IMP::atom::create_protein() to create a coarse grained protein. The function is crude so far, but the idea is that we should have standard ways of generating such things so that everyone can benefit from best practices. - a class IMP::atom::BondGraph has been added so Boost.Graph algorithms can be used with bond graphs - the IMP-specific doxygen commands have been documented. - the doxygen commands unstable{name} and untested{name} have been added to doxygen to be used for marking methods and functions as likely to change or not well tested. Pass the name of the class or method as name. - bonds can now be added to molecular hierarchies automatically see IMP::atom::add_bonds() ## Week of May 1 - There is a IMP::Rotation2D - You can now extract fixed Euler ZYZ angles from a IMP::Rotation3D. This code can serve as a template for extracting other Euler angle conventions if they are needed. - IMP::algebra::rotation_between_two_vectors() changed to IMP::algebra::rotation_taking_first_to_second() to make it clear which direction the rotation goes. ## Week of April 27 - IMP::Decorator derived objects are now convertible to bool. So code like \code if (d) { blah } \endcode works. - IMP::Decorator::cast now returns a null decorator (\c DecoratorName()) rather than throwing an exception. This might make it actually useful. - New IMP::statistics module for statistical methods (e.g. clustering). ## Week of April 20 - Support for taking advantage of rigid bodies in collision detection has been added via the IMP::core::RigidClosePairsFinder. - All IMP::core::ClosePairsScoreState-related functionality has been optimized and should be noticeably faster. - The EMBED project has been merged into the IMP::em module. Thus, it is no longer necessary to obtain EMBED separately to build IMP::em. - All C++ methods which work with std::istream or std::ostream arguments should now be usable from Python: any Python file-like object can be used. - The SWIG interface now ties in to the reference counting mechanism used in the C++ code, for all classes derived from IMP::RefCounted. So, for %example, deleting a Python object should not free the underlying C++ object if another C++ object holds a reference to it, and unrefing a C++ object should not delete it if a Python object still holds a reference to it. There should be no further need to manually keep references to Python objects to prevent things from being prematurely cleaned up. This should be transparent to the user; please report any problems. - IMP::atom::AtomDecorator no longer inherits from IMP::core::XYZDecorator. ## Week of April 13 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator, IMP::atom::ResidueDecorator etc. now inherit from IMP::atom::MolecularHierarchyDecorator. This means they can set their IMP::atom::MolecularHierarchyDecorator::Type properly and saves a lot of typing in the common case. We can always make hierarchy decorators zero memory overhead if we find that the cost matters. - The base classes IMP::Comparable, IMP::NullDefault, IMP::ValidDefault and IMP::UninitializedDefault were added to document what the default constructors do for various objects in IMP. ## Week of April 6 - IMP::Object now defines pure virtual methods IMP::Object::get_version_info() and IMP::Object::show(). All non-abstract classes which inherit from IMP::Object must implement them. Doing this means they are more uniformly implemented and are documented in one place. - IMP::Object is now reference counted (the only objects this changes are IMP::Model and IMP::Optimizer). As a result optimizers and IMP::Model objects must be created on the heap. - IMP::RefCounted has had its name changed and no longer inherits from IMP::Object. This makes it easier to use independently. - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator have been improved - running doxygen produces no warnings. Let's try to keep it that way. Run \command{scons doc > /dev/null} before you check in large blocks of new code to make sure that you didn't introduce warnings. - Helper classes have been added to make SingletonModifier and PairModifier objects look like functors for C++ and Python. This makes it so you can use \c map in Python and \c for_each in C++. See IMP::SingletonFunctor and IMP::PairFunctor for more information. - The recent changes page has moved to a doxygen page to make it easier to link changes back to the relevant source code. ## Week of March 20: - create_rigid_body(), create_rigid_bodies(), create_centroid() etc. now require you to add the score state to the model yourself. Whether the rigid body is enforced or not can be controlled by adding or removing the returned IMP::ScoreState. - All previously deprecated code has been removed. This includes indexes in particles, the no-argument constructor for particles and the functionality that was moved from core to atom. - The IMP::core::RigidBodyDecorator and IMP::core::RigidMemberDecorator have been simplified by removing the traits classes. There is now only one type of rigid body and it stores the list of member particles internally. - IMP::core::NameDecorator has been deprecated as IMP::Particles have built in names (use IMP::Particle::get_name(), IMP::Particle::set_name()). - IMP::atom::ChainDecorator has been added ## Week of March 13: - A new IMP::display::Writer, IMP::display::ChimeraWriter has been added which exports geometry to native Chimera constructs. It probably should be used instead of bild and cmm for most purposes as it combines the best of both types. - compose() for transformation is now a free function so compose(rotation, vector) will work via promotions. - the geometry classes have been cleaned up and are now more consistent with IMP standards (and more likely to do what their names/comments suggest). See IMP::algebra::Sphere3D, IMP::algebra::Cylinder3D, IMP::algebra::Segment3D, IMP::algebra::Plane3D etc. - ParticleRefiner has been renamed to IMP::Refiner since the Particle was long and unnecessary. €eIF1•eIF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3). - IMP::domino::DiscreteSampler::set_restraints() now allows the set of restraints used in DOMINO filtering to be controlled (previously, all restraints in the IMP::Model were used). - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the MODELLER score, not the implicit (and deprecated) score of the entire IMP::Model. - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently treated (they are supposed to be the centers of the voxels, but in some cases they were the lower left edge). The fix may result in maps and their bounding boxes appearing shifted by half the voxel size. # 2.4.0 - 2015-02-27 # {#changelog_2_4_0} - Python 3 is now supported. The Windows package now works out of the box with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu and Fedora. (Apple does not yet provide Python 3 so the Mac package is not built with Python 3 support, but if you install Python 3 you can compile IMP from source to work with it.) - A new command line tool, `rrt_sample`, allows for running rapidly exploring random tree (RRT) based sampling, e.g. to sample conformations of multiple domains connected by flexible linkers. - Command line tools are now placed in the `bin` directory of modules (the old `bin` directory, used for utility programs that are not installed, is now called `utility`). The `applications` directory is removed. - Several Python modules that were only used by EMageFit have been moved from IMP::em2d to IMP::EMageFit. - A new tutorial that uses the IMP::pmi module is now available (modeling of the RNA Polymerase II stalk). # 2.3.1 - 2014-12-01 # {#changelog_2_3_1} - Minor bug fixes. - Update to IMP::pmi to support the Nup84 and rnapolii biological systems. # 2.3.0 - 2014-10-24 # {#changelog_2_3_0} - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental. - Binary installers for 64-bit Windows are now available. - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications. - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used. - The integrative docking (idock) application now employs SOAP to improve the scoring. - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments. - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python. # 2.2.1 - 2014-08-11 # {#changelog_2_2_1} - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/) # 2.2.0 - 2014-04-09 # {#changelog_2_2_0} - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that. - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back. - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry(). - The IMP::base::Flag class was added to simplify adding flags in C++. - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers. - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones. - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls. - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application. - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead. - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added. - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available. - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples. - The container implementations has been cleaned up and simplified a bit. This should not change most usage. - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`). - [Eigen](#Eigen) is now included with IMP. - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead. - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose. - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer. - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h. - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string. - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/. # 2.1.1 - 2013-11-04 # {#changelog_2_1_1} - Added support for Mac OS 10.9 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0} - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time. - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it. - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member. - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before). - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc. - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput. - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation. - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before. - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs. - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages. - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`). - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on. - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested. - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up. - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized. # For IMP 2.0 and earlier # {#changelog_2_0} ## Spring 2013 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned. - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module. - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory. - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS. ## Winter 2012-2013 - Scons support was removed. Use cmake to build IMP, it is way better. - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long. - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody. - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols. - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo. - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems. - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake). - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit. - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file. - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice. - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript. - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case. - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed. - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info. - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far. - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin. - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code. - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page. - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python. - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated. - IMP::base::FailureHandler has been removed as it wasn't used. - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer. - Modules now must be in a folder called `modules`, which must have a `description.py` file. - The build system should now be a little better about cleaning up after itself. - How headers/doc files/data/examples - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name. - Nice version numbers will now be generated if you use git. You need to rerun tools/git/developer_tools/init_git if you have an existing clone. ## Fall 2012 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed. - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro. - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively. - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason. - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know. - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting. - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles. - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container. - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this. - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes. ## Summer 2012 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro. - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API. - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons. - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors. - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling. - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names. ## Winter 2011/2012 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy. - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()). - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes. - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores. - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input). - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use. - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented. - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed. - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects. - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct. - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole. - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed. - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer). - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions. - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified. - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated. - 24/1: \c tools/make-source has been added that adds a source/header pair to a module - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD. - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use. - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations. - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well. - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided) - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values. ## Fall 2011 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector). - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector. - IMP::statistics::HistogramD now supports display with \c matplotlib. - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply. - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn. - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom. - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel. - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d() - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support. - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility. ## Summer 2011 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy. - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone. - IMP::core::ExcludedVolumeRestraint now supports filters - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before. - you can now pick which Python to use with the \c python argument to scons. - the units for BD diffusion coefficients changed to be consistent with other IMP units. - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb. - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform. - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types. - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably. - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated. - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code. - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule. - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step. - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it. - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors. - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants. - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster. - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values. - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference(). - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially faster when used with IMP::Model::evaluate_if_good() - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value. - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder. - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent. - the RMF helper binaries have been moved to their own application so that they get installed. - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written. ## Spring 2011 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster. - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute domino on a cluster as well as debugging the sampling process. - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf. - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz. - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints. - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support. - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them. - The Python graph interface now has a remove_vertex method. - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState. - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier. - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes. ## Winter 2011 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step() - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing. - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful). - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD. - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score. - each IMP::Restraint now has its own weight. This is lightly tested. - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward. - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested). - the name of the clustering methods have been changed to create_ from get_ since they create objects. - the names of the recently added metric based clustering support has been changed to Metric rather than Distance. - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript. - the fuction IMP::display::create_restraint_geometry() was added to do what its name says. - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches. - IMP::atom::DopePairScore has been added to IMP. - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified interface to the bonded part of the CHARMM forcefield. - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons. - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide. - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and residue naming to PDB-style, so CHARMM topology files can be used unmodified. - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes. - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information. - IMP::atom::CHARMMTopology::add_missing_atoms() and IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms read from a PDB file match the CHARMM topology. - building of deprecated functionality is now false by default - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP. - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath). - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead. - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles. - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP. - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is. - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call. - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists. - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info. - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value. - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards. - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this means is that they need to be created with \c new in C++ code. It also means that one is no longer prohibited from storing them (not that there is much use in doing that). - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things. - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better. - the transform based methods on reference frame which have been deprecated for a while were removed - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source. - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version ## Fall 2010 - IMP::display::PlaneGeometry was added to display planes - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to benefit from better checking. - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing. - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer() - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState. - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD. - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory. - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that. - Logging has been improved to better report the name of methods producing the log messages - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output. - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric. - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest. - applications and biological systems now have their own documentation pages - IMP::atom::Hierarchy requires radii on its leaves - IMP::atom::read_pdb add radii to the atoms - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead. - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too. - a new example was added in the kernel to show IMP used on a biological system - IMP::display::PymolWriter now supports adding labels to things - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints. - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions. - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members. - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring. - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls ## Summer 2010 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work. - Javi added his 2D EM scoring functionality as IMP::em2D - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually. - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer). - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc. - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data). - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process - a better explanation of how to use reference counted pointers has been added to IMP::Pointer. - all simple, planar polygons are now supported for display when CGAL is present ## Week of April 1, 2010 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are - types are not checked until the lists are passed to C++ - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python) - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used. - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work. - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying. ## Week of March 13, 2010 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere. - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors). ## Week of March 6, 2010 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite. ## Week of Feb 27, 2010 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis(). - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things - renamed IMP::core::root() to IMP::core::get_root() ## Week of Feb 21, 2010 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means. - various more interpolation and endian methods had their names fixed (by adding \c get_) - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid() - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d() - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions. - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>). - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points. Some good sed strings are - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_in_box(\#get_random_vector_in(\#g" - "s\#random_vector_on_box(\#get_random_vector_on(\#g" - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g" - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g" - "s\#grid_cover\#get_grid_surface_cover\#g" - "s\#uniform_coverr\#get_uniform_surface_cover\#g" . - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD. - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments. - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis() - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point) - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base. - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance() The following global replacements are probably safe (in sed expressions) - "s\#core.distance(\#core.get_distance\#g" - "s\#algebra.distance(\#algebra.get_distance\#g" - "s\#\.transform(\#.get_transformed(\#g" Only in C++ - "s\#\.rotate(\#.get_rotated(\#g" Only in C++ - "s\#rotation_from\#get_rotation_from\#g" - "s\#rotation_in\#get_rotation_in\#g" - "s\#identity_rotation(\#get_identity_rotation_3d(" - "s\#::distance(\#::get_distance(\#g" . - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry. - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future. - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair() ## Week of Feb 14, 2010 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support. - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged. - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use. - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector. - IMP::example renamed (from examples) - The IMP_USAGE_CHECK() macro lost the (unused) third argument. ## Week of February 3, 2010 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model() - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically. - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object. - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show(). - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin. - IMP.atom.create_protein() now returns the created protein - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone(). - IMP.helper.clone() has been renamed to IMP.atom.create_clone() - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner. - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added. - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages. - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement. - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used. - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume(). ## Week of January 22, 2010 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(), IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds. - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer. - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles. - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint. - Add IMP::core::MinumumPairScore (and kin). ## Week of January 1, 2010 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed. - IMP::UsageException and IMP::InternalException are now reserved for their respective checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API exceptions (which are thrown in any build mode and should be documented and tested) and helper exceptions which are only thrown when checks are turned on (and hence should not be documented) clearer. - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves. This means that incremental evaluation of restraints in restraint sets now works and that restraints in restraint sets can be individually evaluated. Existing working usage should not have changed (so \c sed should manage updates just fine), but the internal implementation is completely new, so pay attention. - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are convertible to C++ streams and implicitly constructable from both streams and file paths. As a result, functions that take them will work when called with a string for a file path, a Python file or a C++ stream. - The IMP.examples module now documents how to use the needed aspects of scons and SWIG. - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested. - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there. - boost program options is now a checked for dependency. - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them. - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented. - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file). - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas. - a new base class IMP::Constraint has been added for constraints. Various other classes have been renamed to be consistent. ## Week of December 11 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference. - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed. - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation. - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles). ## Week of December 4 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems. - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data. - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more). ## Week of November 19 - imp.xcodeproj was moved into tools and updated accordingly - the bin directory finally went away - IMP::display::ChimeraWriter now supports the latest version of Chimera - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO. - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method. - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors) ## Week of November 12 - All exception types are now mapped to equivalent classes in Python. The class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception in both C++ and Python) so all IMP exceptions can be caught in Python using "except IMP.Exception". For convenience, the IMP::IOException, IMP::ValueException and IMP::IndexException classes in Python also derive from the similarly-named standard Python classes (IOError, ValueError and IndexError respectively). ## week of October 29 Update - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested). - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box(). - Added IMP::core::VolumeRestraint. - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information. - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy(). - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite. - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs. - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such. - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were. - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added. - All containers now keep track of what they looked like on the previous time step. - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier. - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications. - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints. ## Week of October 15 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement. ## Week of October 8 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems. - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this. - IMP_check() and IMP_assert() have been replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK() respectively. Hopefully this makes it clearer when to use which. IMP::set_check_level() has been revised accordingly (to use IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and IMP_not_implemented() have been renamed to IMP_FAILURE() and IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update external code, - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL - replace IMP.CHEAP with IMP.USAGE - replace IMP_assert and IMP_check as above. ## Week of October 1 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used. - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them. - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly. - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too. ## Week of September 27 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs. - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies. - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking. - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins. ## Week of September 20 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier. - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof. - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always. - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version. ## Week of August 30 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles in a box. - str should now work better in Python when applied to IMP::Object objects and a few others. Objects should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show functions directory to make it so that they work in Python. It is now a bug to \c print or \c str to not return something useful in Python for any IMP object. ## Week of August 23 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur. - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names. - Python documentation is now parsed by doxygen ## Week of August 15 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from. ## Week of August 8 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue. - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements). - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that. - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation. - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency. - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate. - The contents of bin are moving to tools since the two directories both contained assorted development tools (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon. - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored). Anyone who implemented a decorator will have to delete the second argument of their show function. ## Week of August 1 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(), IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(), IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply(). - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle() and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which updates the names in all code in subdirectories of the current directory. Please backup things first. - a function, IMP::Restraint::safe_evaluate() has been added to be used when requesting the value of restraints outside of the IMP::Model::evaluate(). ## Week of July 18 - The atom::BondPairContainer has been renamed to atom::BondPairFilter since it wasn't a real container. A real atom::BondPairContainer has been added. - Out of source builds mostly works. To do an out of source build, make a link to \c scons_tools and \c SConstruct from the source directory in the build directory and then add the line \command{repository="../path_to_src_dir"} in the \c config.py in the build directory. Then running \c scons in the build directory will put generated files there. This allows you to make both a \c release and \c fast build off of the same set of sources. Proper docs will be added soon. - The SConscript setup for a module has been greatly simplified so it should be easy enough for users to maintain their own module and updates to the build system shouldn't involve changes to user SConscripts any more. This does mean that any non-svn modules need to be updated. Look at IMP.core or another module and copy what is there. Again, proper docs will be added soon. ## Week of July 11 - Per Javi's suggestion, the conventions have been split into "Coding conventions" and "API conventions" and linked from the main page. - benchmarks got moved to their own module, IMP::benchmark and will now be tracked over time ## Week of July 4 - FilteredListSingletonContainer and FilteredPairSingletonContainer went away as there was too much hard to factor code shared with IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The IMP::core::ClosePairsScoreState is now handled by a list of filters on the class itself. That is, do IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a filter. - Hooks have been added to allow actions to be taken on an assertion or check failure. See IMP::HandleFailure and its descendants for more details. - IMP::display::CGOWriter has been much improved and now allows provides lots of display-time control and easy building of animations. - Generic optimizer states have been added (IMP::core::SingletonsOptimizerState and IMP::core::PairsOptimizerState). ## Week of June 15 - box distributions (IMP::algebra::random_vector_on_box(), IMP::algebra::random_vector_in_box()) take bounding boxes - IMP::em::DensityMap are now IMP::Object objects. Some current allocations on the stack may need to be changed (since ref counted objects can't be stack allocated). - IMP::em::DensityMap::Read() is now deprecated (as is IMP::em::DensityMap::Write) as they violated a slew of conventions. Use IMP::em::read_map() and IMP::em::write_map() instead. ## Week of June 8 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds(); - Hierarchys to IMP::atom::Hierarchies - IMP::Particles has been changed to a new type. It no longer plays quite so well with Python lists (you can't create one from an arbitrary Python list). - The containers for multiple decorators have been revised to make them more useful. See IMP::Decorators. You can now implicitly convert, for %example, a IMP::core::XYZs to an IMP::Particles (as well as use IMP::Decorators::get_particles()). All methods that take a collection of particles with a certain type now instead take a IMP::Decorators. - A method that has expectations about the particles it takes should take decorators rather than particles. That is, anything which expects IMP::core::XYZ particles, should take an IMP::core::XYZs instead of IMP::Particles. - IMP_LIST() now takes an extra argument (the name of the plural container) - In C++ IMP::Decorator objects are implicitly convertible to IMP::Particle* objects. Unfortunately, making this work in Python is problematic. - More methods take and return decorators rather than IMP::Particles. This means it is clearer what their expectations are and what they produce. ## Week of May 15 - The IMP::helper module was added and various functionality moved there. See its documentation page for a list of what is there. - IMP::algebra::SphericalCoords was renamed to IMP::algebra::SphericalVector3D to be consistent with other algebra classes. - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent with the other similar classes. - We are trying to be a bit more specific about what is expected from a IMP::atom::Hierarchy. Towards that end, there are two methods for extracting a representation from a hierarchy, IMP::atom::get_detailed_representation(), and IMP::atom::get_simplified_representation() - A method has been added to produce a simplified copy of an IMP::atom::Hierarchy. The method is still a bit crude and could benefit from other people inspecting the output and saying how it should be improved. - There is now an IMP::algebra::BoundingBoxD (and IMP::algebra::BoundingBox3D) to track bounding boxes. There are methods like IMP::algebra::bound_box(Sphere3D) which gets the bounding box of a sphere. - Each IMP::Object has its own log level. If IMP::Object::set_log_level() is called with a value other than IMP::DEFAULT, the value is used for code within and called by the object. To support this in your objects you need to add IMP_OBJECT_LOG to the start of each major method. It has been added to some of the core objects. ## Week of May 8 - IMP::atom::MolecularHierarchy has been renamed to IMP::atom::Hierarchy. Use \command{sed -i.old "s\#MolecularHierarchy\#Hierarchy\#g" my_files} - the long-deprecated Name(Decorator) was removed - The decorator suffix has been removed from decorator names. - Unit conventions have been established and documented for distance, energy, force and angle. - there is a function IMP::atom::create_protein() to create a coarse grained protein. The function is crude so far, but the idea is that we should have standard ways of generating such things so that everyone can benefit from best practices. - a class IMP::atom::BondGraph has been added so Boost.Graph algorithms can be used with bond graphs - the IMP-specific doxygen commands have been documented. - the doxygen commands unstable{name} and untested{name} have been added to doxygen to be used for marking methods and functions as likely to change or not well tested. Pass the name of the class or method as name. - bonds can now be added to molecular hierarchies automatically see IMP::atom::add_bonds() ## Week of May 1 - There is a IMP::Rotation2D - You can now extract fixed Euler ZYZ angles from a IMP::Rotation3D. This code can serve as a template for extracting other Euler angle conventions if they are needed. - IMP::algebra::rotation_between_two_vectors() changed to IMP::algebra::rotation_taking_first_to_second() to make it clear which direction the rotation goes. ## Week of April 27 - IMP::Decorator derived objects are now convertible to bool. So code like \code if (d) { blah } \endcode works. - IMP::Decorator::cast now returns a null decorator (\c DecoratorName()) rather than throwing an exception. This might make it actually useful. - New IMP::statistics module for statistical methods (e.g. clustering). ## Week of April 20 - Support for taking advantage of rigid bodies in collision detection has been added via the IMP::core::RigidClosePairsFinder. - All IMP::core::ClosePairsScoreState-related functionality has been optimized and should be noticeably faster. - The EMBED project has been merged into the IMP::em module. Thus, it is no longer necessary to obtain EMBED separately to build IMP::em. - All C++ methods which work with std::istream or std::ostream arguments should now be usable from Python: any Python file-like object can be used. - The SWIG interface now ties in to the reference counting mechanism used in the C++ code, for all classes derived from IMP::RefCounted. So, for %example, deleting a Python object should not free the underlying C++ object if another C++ object holds a reference to it, and unrefing a C++ object should not delete it if a Python object still holds a reference to it. There should be no further need to manually keep references to Python objects to prevent things from being prematurely cleaned up. This should be transparent to the user; please report any problems. - IMP::atom::AtomDecorator no longer inherits from IMP::core::XYZDecorator. ## Week of April 13 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator, IMP::atom::ResidueDecorator etc. now inherit from IMP::atom::MolecularHierarchyDecorator. This means they can set their IMP::atom::MolecularHierarchyDecorator::Type properly and saves a lot of typing in the common case. We can always make hierarchy decorators zero memory overhead if we find that the cost matters. - The base classes IMP::Comparable, IMP::NullDefault, IMP::ValidDefault and IMP::UninitializedDefault were added to document what the default constructors do for various objects in IMP. ## Week of April 6 - IMP::Object now defines pure virtual methods IMP::Object::get_version_info() and IMP::Object::show(). All non-abstract classes which inherit from IMP::Object must implement them. Doing this means they are more uniformly implemented and are documented in one place. - IMP::Object is now reference counted (the only objects this changes are IMP::Model and IMP::Optimizer). As a result optimizers and IMP::Model objects must be created on the heap. - IMP::RefCounted has had its name changed and no longer inherits from IMP::Object. This makes it easier to use independently. - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator have been improved - running doxygen produces no warnings. Let's try to keep it that way. Run \command{scons doc > /dev/null} before you check in large blocks of new code to make sure that you didn't introduce warnings. - Helper classes have been added to make SingletonModifier and PairModifier objects look like functors for C++ and Python. This makes it so you can use \c map in Python and \c for_each in C++. See IMP::SingletonFunctor and IMP::PairFunctor for more information. - The recent changes page has moved to a doxygen page to make it easier to link changes back to the relevant source code. ## Week of March 20: - create_rigid_body(), create_rigid_bodies(), create_centroid() etc. now require you to add the score state to the model yourself. Whether the rigid body is enforced or not can be controlled by adding or removing the returned IMP::ScoreState. - All previously deprecated code has been removed. This includes indexes in particles, the no-argument constructor for particles and the functionality that was moved from core to atom. - The IMP::core::RigidBodyDecorator and IMP::core::RigidMemberDecorator have been simplified by removing the traits classes. There is now only one type of rigid body and it stores the list of member particles internally. - IMP::core::NameDecorator has been deprecated as IMP::Particles have built in names (use IMP::Particle::get_name(), IMP::Particle::set_name()). - IMP::atom::ChainDecorator has been added ## Week of March 13: - A new IMP::display::Writer, IMP::display::ChimeraWriter has been added which exports geometry to native Chimera constructs. It probably should be used instead of bild and cmm for most purposes as it combines the best of both types. - compose() for transformation is now a free function so compose(rotation, vector) will work via promotions. - the geometry classes have been cleaned up and are now more consistent with IMP standards (and more likely to do what their names/comments suggest). See IMP::algebra::Sphere3D, IMP::algebra::Cylinder3D, IMP::algebra::Segment3D, IMP::algebra::Plane3D etc. - ParticleRefiner has been renamed to IMP::Refiner since the Particle was long and unnecessary. ˘IF1•eIF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3). - IMP::domino::DiscreteSampler::set_restraints() now allows the set of restraints used in DOMINO filtering to be controlled (previously, all restraints in the IMP::Model were used). - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the MODELLER score, not the implicit (and deprecated) score of the entire IMP::Model. - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently treated (they are supposed to be the centers of the voxels, but in some cases they were the lower left edge). The fix may result in maps and their bounding boxes appearing shifted by half the voxel size. # 2.4.0 - 2015-02-27 # {#changelog_2_4_0} - Python 3 is now supported. The Windows package now works out of the box with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu and Fedora. (Apple does not yet provide Python 3 so the Mac package is not built with Python 3 support, but if you install Python 3 you can compile IMP from source to work with it.) - A new command line tool, `rrt_sample`, allows for running rapidly exploring random tree (RRT) based sampling, e.g. to sample conformations of multiple domains connected by flexible linkers. - Command line tools are now placed in the `bin` directory of modules (the old `bin` directory, used for utility programs that are not installed, is now called `utility`). The `applications` directory is removed. - Several Python modules that were only used by EMageFit have been moved from IMP::em2d to IMP::EMageFit. - A new tutorial that uses the IMP::pmi module is now available (modeling of the RNA Polymerase II stalk). # 2.3.1 - 2014-12-01 # {#changelog_2_3_1} - Minor bug fixes. - Update to IMP::pmi to support the Nup84 and rnapolii biological systems. # 2.3.0 - 2014-10-24 # {#changelog_2_3_0} - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental. - Binary installers for 64-bit Windows are now available. - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications. - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used. - The integrative docking (idock) application now employs SOAP to improve the scoring. - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments. - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python. # 2.2.1 - 2014-08-11 # {#changelog_2_2_1} - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/) # 2.2.0 - 2014-04-09 # {#changelog_2_2_0} - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that. - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back. - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry(). - The IMP::base::Flag class was added to simplify adding flags in C++. - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers. - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones. - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls. - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application. - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead. - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added. - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available. - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples. - The container implementations has been cleaned up and simplified a bit. This should not change most usage. - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`). - [Eigen](#Eigen) is now included with IMP. - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead. - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose. - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer. - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h. - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string. - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/. # 2.1.1 - 2013-11-04 # {#changelog_2_1_1} - Added support for Mac OS 10.9 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0} - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time. - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it. - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member. - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before). - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc. - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput. - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation. - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before. - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs. - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages. - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`). - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on. - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested. - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up. - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized. # For IMP 2.0 and earlier # {#changelog_2_0} ## Spring 2013 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned. - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module. - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory. - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS. ## Winter 2012-2013 - Scons support was removed. Use cmake to build IMP, it is way better. - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long. - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody. - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols. - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo. - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems. - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake). - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit. - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file. - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice. - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript. - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case. - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed. - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info. - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far. - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin. - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code. - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page. - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python. - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated. - IMP::base::FailureHandler has been removed as it wasn't used. - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer. - Modules now must be in a folder called `modules`, which must have a `description.py` file. - The build system should now be a little better about cleaning up after itself. - How headers/doc files/data/examples - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name. - Nice version numbers will now be generated if you use git. You need to rerun tools/git/developer_tools/init_git if you have an existing clone. ## Fall 2012 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed. - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro. - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively. - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason. - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know. - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting. - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles. - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container. - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this. - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes. ## Summer 2012 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro. - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API. - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons. - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors. - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling. - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names. ## Winter 2011/2012 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy. - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()). - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes. - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores. - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input). - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use. - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented. - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed. - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects. - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct. - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole. - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed. - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer). - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions. - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified. - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated. - 24/1: \c tools/make-source has been added that adds a source/header pair to a module - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD. - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use. - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations. - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well. - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided) - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values. ## Fall 2011 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector). - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector. - IMP::statistics::HistogramD now supports display with \c matplotlib. - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply. - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn. - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom. - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel. - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d() - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support. - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility. ## Summer 2011 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy. - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone. - IMP::core::ExcludedVolumeRestraint now supports filters - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before. - you can now pick which Python to use with the \c python argument to scons. - the units for BD diffusion coefficients changed to be consistent with other IMP units. - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb. - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform. - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types. - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably. - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated. - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code. - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule. - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step. - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it. - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors. - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants. - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster. - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values. - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference(). - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially faster when used with IMP::Model::evaluate_if_good() - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value. - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder. - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent. - the RMF helper binaries have been moved to their own application so that they get installed. - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written. ## Spring 2011 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster. - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute domino on a cluster as well as debugging the sampling process. - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf. - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz. - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints. - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support. - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them. - The Python graph interface now has a remove_vertex method. - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState. - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier. - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes. ## Winter 2011 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step() - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing. - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful). - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD. - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score. - each IMP::Restraint now has its own weight. This is lightly tested. - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward. - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested). - the name of the clustering methods have been changed to create_ from get_ since they create objects. - the names of the recently added metric based clustering support has been changed to Metric rather than Distance. - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript. - the fuction IMP::display::create_restraint_geometry() was added to do what its name says. - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches. - IMP::atom::DopePairScore has been added to IMP. - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified interface to the bonded part of the CHARMM forcefield. - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons. - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide. - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and residue naming to PDB-style, so CHARMM topology files can be used unmodified. - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes. - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information. - IMP::atom::CHARMMTopology::add_missing_atoms() and IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms read from a PDB file match the CHARMM topology. - building of deprecated functionality is now false by default - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP. - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath). - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead. - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles. - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP. - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is. - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call. - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists. - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info. - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value. - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards. - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this means is that they need to be created with \c new in C++ code. It also means that one is no longer prohibited from storing them (not that there is much use in doing that). - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things. - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better. - the transform based methods on reference frame which have been deprecated for a while were removed - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source. - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version ## Fall 2010 - IMP::display::PlaneGeometry was added to display planes - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to benefit from better checking. - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing. - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer() - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState. - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD. - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory. - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that. - Logging has been improved to better report the name of methods producing the log messages - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output. - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric. - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest. - applications and biological systems now have their own documentation pages - IMP::atom::Hierarchy requires radii on its leaves - IMP::atom::read_pdb add radii to the atoms - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead. - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too. - a new example was added in the kernel to show IMP used on a biological system - IMP::display::PymolWriter now supports adding labels to things - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints. - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions. - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members. - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring. - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls ## Summer 2010 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work. - Javi added his 2D EM scoring functionality as IMP::em2D - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually. - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer). - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc. - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data). - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process - a better explanation of how to use reference counted pointers has been added to IMP::Pointer. - all simple, planar polygons are now supported for display when CGAL is present ## Week of April 1, 2010 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are - types are not checked until the lists are passed to C++ - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python) - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used. - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work. - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying. ## Week of March 13, 2010 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere. - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors). ## Week of March 6, 2010 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite. ## Week of Feb 27, 2010 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis(). - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things - renamed IMP::core::root() to IMP::core::get_root() ## Week of Feb 21, 2010 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means. - various more interpolation and endian methods had their names fixed (by adding \c get_) - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid() - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d() - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions. - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>). - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points. Some good sed strings are - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_in_box(\#get_random_vector_in(\#g" - "s\#random_vector_on_box(\#get_random_vector_on(\#g" - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g" - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g" - "s\#grid_cover\#get_grid_surface_cover\#g" - "s\#uniform_coverr\#get_uniform_surface_cover\#g" . - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD. - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments. - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis() - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point) - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base. - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance() The following global replacements are probably safe (in sed expressions) - "s\#core.distance(\#core.get_distance\#g" - "s\#algebra.distance(\#algebra.get_distance\#g" - "s\#\.transform(\#.get_transformed(\#g" Only in C++ - "s\#\.rotate(\#.get_rotated(\#g" Only in C++ - "s\#rotation_from\#get_rotation_from\#g" - "s\#rotation_in\#get_rotation_in\#g" - "s\#identity_rotation(\#get_identity_rotation_3d(" - "s\#::distance(\#::get_distance(\#g" . - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry. - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future. - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair() ## Week of Feb 14, 2010 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support. - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged. - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use. - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector. - IMP::example renamed (from examples) - The IMP_USAGE_CHECK() macro lost the (unused) third argument. ## Week of February 3, 2010 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model() - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically. - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object. - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show(). - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin. - IMP.atom.create_protein() now returns the created protein - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone(). - IMP.helper.clone() has been renamed to IMP.atom.create_clone() - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner. - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added. - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages. - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement. - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used. - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume(). ## Week of January 22, 2010 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(), IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds. - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer. - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles. - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint. - Add IMP::core::MinumumPairScore (and kin). ## Week of January 1, 2010 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed. - IMP::UsageException and IMP::InternalException are now reserved for their respective checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API exceptions (which are thrown in any build mode and should be documented and tested) and helper exceptions which are only thrown when checks are turned on (and hence should not be documented) clearer. - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves. This means that incremental evaluation of restraints in restraint sets now works and that restraints in restraint sets can be individually evaluated. Existing working usage should not have changed (so \c sed should manage updates just fine), but the internal implementation is completely new, so pay attention. - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are convertible to C++ streams and implicitly constructable from both streams and file paths. As a result, functions that take them will work when called with a string for a file path, a Python file or a C++ stream. - The IMP.examples module now documents how to use the needed aspects of scons and SWIG. - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested. - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there. - boost program options is now a checked for dependency. - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them. - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented. - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file). - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas. - a new base class IMP::Constraint has been added for constraints. Various other classes have been renamed to be consistent. ## Week of December 11 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference. - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed. - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation. - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles). ## Week of December 4 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems. - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data. - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more). ## Week of November 19 - imp.xcodeproj was moved into tools and updated accordingly - the bin directory finally went away - IMP::display::ChimeraWriter now supports the latest version of Chimera - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO. - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method. - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors) ## Week of November 12 - All exception types are now mapped to equivalent classes in Python. The class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception in both C++ and Python) so all IMP exceptions can be caught in Python using "except IMP.Exception". For convenience, the IMP::IOException, IMP::ValueException and IMP::IndexException classes in Python also derive from the similarly-named standard Python classes (IOError, ValueError and IndexError respectively). ## week of October 29 Update - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested). - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box(). - Added IMP::core::VolumeRestraint. - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information. - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy(). - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite. - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs. - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such. - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were. - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added. - All containers now keep track of what they looked like on the previous time step. - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier. - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications. - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints. ## Week of October 15 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement. ## Week of October 8 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems. - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this. - IMP_check() and IMP_assert() have been replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK() respectively. Hopefully this makes it clearer when to use which. IMP::set_check_level() has been revised accordingly (to use IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and IMP_not_implemented() have been renamed to IMP_FAILURE() and IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update external code, - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL - replace IMP.CHEAP with IMP.USAGE - replace IMP_assert and IMP_check as above. ## Week of October 1 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used. - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them. - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly. - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too. ## Week of September 27 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs. - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies. - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking. - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins. ## Week of September 20 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier. - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof. - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always. - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version. ## Week of August 30 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles in a box. - str should now work better in Python when applied to IMP::Object objects and a few others. Objects should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show functions directory to make it so that they work in Python. It is now a bug to \c print or \c str to not return something useful in Python for any IMP object. ## Week of August 23 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur. - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names. - Python documentation is now parsed by doxygen ## Week of August 15 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from. ## Week of August 8 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue. - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements). - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that. - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation. - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency. - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate. - The contents of bin are moving to tools since the two directories both contained assorted development tools (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon. - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored). Anyone who implemented a decorator will have to delete the second argument of their show function. ## Week of August 1 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(), IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(), IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply(). - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle() and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which updates the names in all code in subdirectories of the current directory. Please backup things first. - a function, IMP::Restraint::safe_evaluate() has been added to be used when requesting the value of restraints outside of the IMP::Model::evaluate(). ## Week of July 18 - The atom::BondPairContainer has been renamed to atom::BondPairFilter since it wasn't a real container. A real atom::BondPairContainer has been added. - Out of source builds mostly works. To do an out of source build, make a link to \c scons_tools and \c SConstruct from the source directory in the build directory and then add the line \command{repository="../path_to_src_dir"} in the \c config.py in the build directory. Then running \c scons in the build directory will put generated files there. This allows you to make both a \c release and \c fast build off of the same set of sources. Proper docs will be added soon. - The SConscript setup for a module has been greatly simplified so it should be easy enough for users to maintain their own module and updates to the build system shouldn't involve changes to user SConscripts any more. This does mean that any non-svn modules need to be updated. Look at IMP.core or another module and copy what is there. Again, proper docs will be added soon. ## Week of July 11 - Per Javi's suggestion, the conventions have been split into "Coding conventions" and "API conventions" and linked from the main page. - benchmarks got moved to their own module, IMP::benchmark and will now be tracked over time ## Week of July 4 - FilteredListSingletonContainer and FilteredPairSingletonContainer went away as there was too much hard to factor code shared with IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The IMP::core::ClosePairsScoreState is now handled by a list of filters on the class itself. That is, do IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a filter. - Hooks have been added to allow actions to be taken on an assertion or check failure. See IMP::HandleFailure and its descendants for more details. - IMP::display::CGOWriter has been much improved and now allows provides lots of display-time control and easy building of animations. - Generic optimizer states have been added (IMP::core::SingletonsOptimizerState and IMP::core::PairsOptimizerState). ## Week of June 15 - box distributions (IMP::algebra::random_vector_on_box(), IMP::algebra::random_vector_in_box()) take bounding boxes - IMP::em::DensityMap are now IMP::Object objects. Some current allocations on the stack may need to be changed (since ref counted objects can't be stack allocated). - IMP::em::DensityMap::Read() is now deprecated (as is IMP::em::DensityMap::Write) as they violated a slew of conventions. Use IMP::em::read_map() and IMP::em::write_map() instead. ## Week of June 8 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds(); - Hierarchys to IMP::atom::Hierarchies - IMP::Particles has been changed to a new type. It no longer plays quite so well with Python lists (you can't create one from an arbitrary Python list). - The containers for multiple decorators have been revised to make them more useful. See IMP::Decorators. You can now implicitly convert, for %example, a IMP::core::XYZs to an IMP::Particles (as well as use IMP::Decorators::get_particles()). All methods that take a collection of particles with a certain type now instead take a IMP::Decorators. - A method that has expectations about the particles it takes should take decorators rather than particles. That is, anything which expects IMP::core::XYZ particles, should take an IMP::core::XYZs instead of IMP::Particles. - IMP_LIST() now takes an extra argument (the name of the plural container) - In C++ IMP::Decorator objects are implicitly convertible to IMP::Particle* objects. Unfortunately, making this work in Python is problematic. - More methods take and return decorators rather than IMP::Particles. This means it is clearer what their expectations are and what they produce. ## Week of May 15 - The IMP::helper module was added and various functionality moved there. See its documentation page for a list of what is there. - IMP::algebra::SphericalCoords was renamed to IMP::algebra::SphericalVector3D to be consistent with other algebra classes. - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent with the other similar classes. - We are trying to be a bit more specific about what is expected from a IMP::atom::Hierarchy. Towards that end, there are two methods for extracting a representation from a hierarchy, IMP::atom::get_detailed_representation(), and IMP::atom::get_simplified_representation() - A method has been added to produce a simplified copy of an IMP::atom::Hierarchy. The method is still a bit crude and could benefit from other people inspecting the output and saying how it should be improved. - There is now an IMP::algebra::BoundingBoxD (and IMP::algebra::BoundingBox3D) to track bounding boxes. There are methods like IMP::algebra::bound_box(Sphere3D) which gets the bounding box of a sphere. - Each IMP::Object has its own log level. If IMP::Object::set_log_level() is called with a value other than IMP::DEFAULT, the value is used for code within and called by the object. To support this in your objects you need to add IMP_OBJECT_LOG to the start of each major method. It has been added to some of the core objects. ## Week of May 8 - IMP::atom::MolecularHierarchy has been renamed to IMP::atom::Hierarchy. Use \command{sed -i.old "s\#MolecularHierarchy\#Hierarchy\#g" my_files} - the long-deprecated Name(Decorator) was removed - The decorator suffix has been removed from decorator names. - Unit conventions have been established and documented for distance, energy, force and angle. - there is a function IMP::atom::create_protein() to create a coarse grained protein. The function is crude so far, but the idea is that we should have standard ways of generating such things so that everyone can benefit from best practices. - a class IMP::atom::BondGraph has been added so Boost.Graph algorithms can be used with bond graphs - the IMP-specific doxygen commands have been documented. - the doxygen commands unstable{name} and untested{name} have been added to doxygen to be used for marking methods and functions as likely to change or not well tested. Pass the name of the class or method as name. - bonds can now be added to molecular hierarchies automatically see IMP::atom::add_bonds() ## Week of May 1 - There is a IMP::Rotation2D - You can now extract fixed Euler ZYZ angles from a IMP::Rotation3D. This code can serve as a template for extracting other Euler angle conventions if they are needed. - IMP::algebra::rotation_between_two_vectors() changed to IMP::algebra::rotation_taking_first_to_second() to make it clear which direction the rotation goes. ## Week of April 27 - IMP::Decorator derived objects are now convertible to bool. So code like \code if (d) { blah } \endcode works. - IMP::Decorator::cast now returns a null decorator (\c DecoratorName()) rather than throwing an exception. This might make it actually useful. - New IMP::statistics module for statistical methods (e.g. clustering). ## Week of April 20 - Support for taking advantage of rigid bodies in collision detection has been added via the IMP::core::RigidClosePairsFinder. - All IMP::core::ClosePairsScoreState-related functionality has been optimized and should be noticeably faster. - The EMBED project has been merged into the IMP::em module. Thus, it is no longer necessary to obtain EMBED separately to build IMP::em. - All C++ methods which work with std::istream or std::ostream arguments should now be usable from Python: any Python file-like object can be used. - The SWIG interface now ties in to the reference counting mechanism used in the C++ code, for all classes derived from IMP::RefCounted. So, for %example, deleting a Python object should not free the underlying C++ object if another C++ object holds a reference to it, and unrefing a C++ object should not delete it if a Python object still holds a reference to it. There should be no further need to manually keep references to Python objects to prevent things from being prematurely cleaned up. This should be transparent to the user; please report any problems. - IMP::atom::AtomDecorator no longer inherits from IMP::core::XYZDecorator. ## Week of April 13 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator, IMP::atom::ResidueDecorator etc. now inherit from IMP::atom::MolecularHierarchyDecorator. This means they can set their IMP::atom::MolecularHierarchyDecorator::Type properly and saves a lot of typing in the common case. We can always make hierarchy decorators zero memory overhead if we find that the cost matters. - The base classes IMP::Comparable, IMP::NullDefault, IMP::ValidDefault and IMP::UninitializedDefault were added to document what the default constructors do for various objects in IMP. ## Week of April 6 - IMP::Object now defines pure virtual methods IMP::Object::get_version_info() and IMP::Object::show(). All non-abstract classes which inherit from IMP::Object must implement them. Doing this means they are more uniformly implemented and are documented in one place. - IMP::Object is now reference counted (the only objects this changes are IMP::Model and IMP::Optimizer). As a result optimizers and IMP::Model objects must be created on the heap. - IMP::RefCounted has had its name changed and no longer inherits from IMP::Object. This makes it easier to use independently. - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator have been improved - running doxygen produces no warnings. Let's try to keep it that way. Run \command{scons doc > /dev/null} before you check in large blocks of new code to make sure that you didn't introduce warnings. - Helper classes have been added to make SingletonModifier and PairModifier objects look like functors for C++ and Python. This makes it so you can use \c map in Python and \c for_each in C++. See IMP::SingletonFunctor and IMP::PairFunctor for more information. - The recent changes page has moved to a doxygen page to make it easier to link changes back to the relevant source code. ## Week of March 20: - create_rigid_body(), create_rigid_bodies(), create_centroid() etc. now require you to add the score state to the model yourself. Whether the rigid body is enforced or not can be controlled by adding or removing the returned IMP::ScoreState. - All previously deprecated code has been removed. This includes indexes in particles, the no-argument constructor for particles and the functionality that was moved from core to atom. - The IMP::core::RigidBodyDecorator and IMP::core::RigidMemberDecorator have been simplified by removing the traits classes. There is now only one type of rigid body and it stores the list of member particles internally. - IMP::core::NameDecorator has been deprecated as IMP::Particles have built in names (use IMP::Particle::get_name(), IMP::Particle::set_name()). - IMP::atom::ChainDecorator has been added ## Week of March 13: - A new IMP::display::Writer, IMP::display::ChimeraWriter has been added which exports geometry to native Chimera constructs. It probably should be used instead of bild and cmm for most purposes as it combines the best of both types. - compose() for transformation is now a free function so compose(rotation, vector) will work via promotions. - the geometry classes have been cleaned up and are now more consistent with IMP standards (and more likely to do what their names/comments suggest). See IMP::algebra::Sphere3D, IMP::algebra::Cylinder3D, IMP::algebra::Segment3D, IMP::algebra::Plane3D etc. - ParticleRefiner has been renamed to IMP::Refiner since the Particle was long and unnecessary. eIF1â˘eIF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3). - IMP::domino::DiscreteSampler::set_restraints() now allows the set of restraints used in DOMINO filtering to be controlled (previously, all restraints in the IMP::Model were used). - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the MODELLER score, not the implicit (and deprecated) score of the entire IMP::Model. - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently treated (they are supposed to be the centers of the voxels, but in some cases they were the lower left edge). The fix may result in maps and their bounding boxes appearing shifted by half the voxel size. # 2.4.0 - 2015-02-27 # {#changelog_2_4_0} - Python 3 is now supported. The Windows package now works out of the box with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu and Fedora. (Apple does not yet provide Python 3 so the Mac package is not built with Python 3 support, but if you install Python 3 you can compile IMP from source to work with it.) - A new command line tool, `rrt_sample`, allows for running rapidly exploring random tree (RRT) based sampling, e.g. to sample conformations of multiple domains connected by flexible linkers. - Command line tools are now placed in the `bin` directory of modules (the old `bin` directory, used for utility programs that are not installed, is now called `utility`). The `applications` directory is removed. - Several Python modules that were only used by EMageFit have been moved from IMP::em2d to IMP::EMageFit. - A new tutorial that uses the IMP::pmi module is now available (modeling of the RNA Polymerase II stalk). # 2.3.1 - 2014-12-01 # {#changelog_2_3_1} - Minor bug fixes. - Update to IMP::pmi to support the Nup84 and rnapolii biological systems. # 2.3.0 - 2014-10-24 # {#changelog_2_3_0} - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental. - Binary installers for 64-bit Windows are now available. - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications. - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used. - The integrative docking (idock) application now employs SOAP to improve the scoring. - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments. - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python. # 2.2.1 - 2014-08-11 # {#changelog_2_2_1} - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/) # 2.2.0 - 2014-04-09 # {#changelog_2_2_0} - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that. - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back. - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry(). - The IMP::base::Flag class was added to simplify adding flags in C++. - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers. - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones. - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls. - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application. - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead. - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added. - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available. - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples. - The container implementations has been cleaned up and simplified a bit. This should not change most usage. - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`). - [Eigen](#Eigen) is now included with IMP. - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead. - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose. - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer. - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h. - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string. - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/. # 2.1.1 - 2013-11-04 # {#changelog_2_1_1} - Added support for Mac OS 10.9 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0} - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time. - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it. - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member. - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before). - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc. - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput. - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation. - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before. - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs. - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages. - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`). - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on. - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested. - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up. - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized. # For IMP 2.0 and earlier # {#changelog_2_0} ## Spring 2013 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned. - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module. - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory. - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS. ## Winter 2012-2013 - Scons support was removed. Use cmake to build IMP, it is way better. - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long. - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody. - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols. - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo. - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems. - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake). - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit. - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file. - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice. - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript. - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case. - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed. - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info. - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far. - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin. - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code. - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page. - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python. - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated. - IMP::base::FailureHandler has been removed as it wasn't used. - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer. - Modules now must be in a folder called `modules`, which must have a `description.py` file. - The build system should now be a little better about cleaning up after itself. - How headers/doc files/data/examples - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name. - Nice version numbers will now be generated if you use git. You need to rerun tools/git/developer_tools/init_git if you have an existing clone. ## Fall 2012 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed. - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro. - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively. - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason. - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know. - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting. - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles. - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container. - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this. - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes. ## Summer 2012 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro. - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API. - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons. - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors. - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling. - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names. ## Winter 2011/2012 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy. - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()). - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes. - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores. - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input). - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use. - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented. - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed. - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects. - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct. - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole. - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed. - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer). - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions. - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified. - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated. - 24/1: \c tools/make-source has been added that adds a source/header pair to a module - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD. - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use. - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations. - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well. - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided) - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values. ## Fall 2011 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector). - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector. - IMP::statistics::HistogramD now supports display with \c matplotlib. - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply. - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn. - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom. - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel. - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d() - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support. - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility. ## Summer 2011 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy. - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone. - IMP::core::ExcludedVolumeRestraint now supports filters - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before. - you can now pick which Python to use with the \c python argument to scons. - the units for BD diffusion coefficients changed to be consistent with other IMP units. - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb. - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform. - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types. - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably. - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated. - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code. - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule. - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step. - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it. - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors. - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants. - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster. - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values. - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference(). - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially faster when used with IMP::Model::evaluate_if_good() - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value. - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder. - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent. - the RMF helper binaries have been moved to their own application so that they get installed. - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written. ## Spring 2011 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster. - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute domino on a cluster as well as debugging the sampling process. - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf. - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz. - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints. - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support. - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them. - The Python graph interface now has a remove_vertex method. - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState. - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier. - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes. ## Winter 2011 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step() - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing. - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful). - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD. - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score. - each IMP::Restraint now has its own weight. This is lightly tested. - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward. - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested). - the name of the clustering methods have been changed to create_ from get_ since they create objects. - the names of the recently added metric based clustering support has been changed to Metric rather than Distance. - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript. - the fuction IMP::display::create_restraint_geometry() was added to do what its name says. - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches. - IMP::atom::DopePairScore has been added to IMP. - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified interface to the bonded part of the CHARMM forcefield. - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons. - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide. - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and residue naming to PDB-style, so CHARMM topology files can be used unmodified. - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes. - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information. - IMP::atom::CHARMMTopology::add_missing_atoms() and IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms read from a PDB file match the CHARMM topology. - building of deprecated functionality is now false by default - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP. - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath). - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead. - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles. - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP. - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is. - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call. - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists. - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info. - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value. - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards. - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this means is that they need to be created with \c new in C++ code. It also means that one is no longer prohibited from storing them (not that there is much use in doing that). - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things. - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better. - the transform based methods on reference frame which have been deprecated for a while were removed - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source. - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version ## Fall 2010 - IMP::display::PlaneGeometry was added to display planes - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to benefit from better checking. - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing. - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer() - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState. - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD. - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory. - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that. - Logging has been improved to better report the name of methods producing the log messages - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output. - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric. - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest. - applications and biological systems now have their own documentation pages - IMP::atom::Hierarchy requires radii on its leaves - IMP::atom::read_pdb add radii to the atoms - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead. - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too. - a new example was added in the kernel to show IMP used on a biological system - IMP::display::PymolWriter now supports adding labels to things - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints. - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions. - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members. - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring. - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls ## Summer 2010 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work. - Javi added his 2D EM scoring functionality as IMP::em2D - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually. - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer). - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc. - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data). - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process - a better explanation of how to use reference counted pointers has been added to IMP::Pointer. - all simple, planar polygons are now supported for display when CGAL is present ## Week of April 1, 2010 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are - types are not checked until the lists are passed to C++ - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python) - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used. - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work. - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying. ## Week of March 13, 2010 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere. - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors). ## Week of March 6, 2010 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite. ## Week of Feb 27, 2010 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis(). - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things - renamed IMP::core::root() to IMP::core::get_root() ## Week of Feb 21, 2010 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means. - various more interpolation and endian methods had their names fixed (by adding \c get_) - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid() - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d() - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions. - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>). - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points. Some good sed strings are - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_in_box(\#get_random_vector_in(\#g" - "s\#random_vector_on_box(\#get_random_vector_on(\#g" - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g" - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g" - "s\#grid_cover\#get_grid_surface_cover\#g" - "s\#uniform_coverr\#get_uniform_surface_cover\#g" . - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD. - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments. - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis() - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point) - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base. - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance() The following global replacements are probably safe (in sed expressions) - "s\#core.distance(\#core.get_distance\#g" - "s\#algebra.distance(\#algebra.get_distance\#g" - "s\#\.transform(\#.get_transformed(\#g" Only in C++ - "s\#\.rotate(\#.get_rotated(\#g" Only in C++ - "s\#rotation_from\#get_rotation_from\#g" - "s\#rotation_in\#get_rotation_in\#g" - "s\#identity_rotation(\#get_identity_rotation_3d(" - "s\#::distance(\#::get_distance(\#g" . - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry. - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future. - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair() ## Week of Feb 14, 2010 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support. - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged. - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use. - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector. - IMP::example renamed (from examples) - The IMP_USAGE_CHECK() macro lost the (unused) third argument. ## Week of February 3, 2010 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model() - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically. - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object. - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show(). - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin. - IMP.atom.create_protein() now returns the created protein - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone(). - IMP.helper.clone() has been renamed to IMP.atom.create_clone() - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner. - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added. - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages. - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement. - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used. - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume(). ## Week of January 22, 2010 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(), IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds. - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer. - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles. - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint. - Add IMP::core::MinumumPairScore (and kin). ## Week of January 1, 2010 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed. - IMP::UsageException and IMP::InternalException are now reserved for their respective checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API exceptions (which are thrown in any build mode and should be documented and tested) and helper exceptions which are only thrown when checks are turned on (and hence should not be documented) clearer. - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves. This means that incremental evaluation of restraints in restraint sets now works and that restraints in restraint sets can be individually evaluated. Existing working usage should not have changed (so \c sed should manage updates just fine), but the internal implementation is completely new, so pay attention. - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are convertible to C++ streams and implicitly constructable from both streams and file paths. As a result, functions that take them will work when called with a string for a file path, a Python file or a C++ stream. - The IMP.examples module now documents how to use the needed aspects of scons and SWIG. - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested. - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there. - boost program options is now a checked for dependency. - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them. - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented. - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file). - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas. - a new base class IMP::Constraint has been added for constraints. Various other classes have been renamed to be consistent. ## Week of December 11 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference. - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed. - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation. - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles). ## Week of December 4 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems. - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data. - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more). ## Week of November 19 - imp.xcodeproj was moved into tools and updated accordingly - the bin directory finally went away - IMP::display::ChimeraWriter now supports the latest version of Chimera - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO. - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method. - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors) ## Week of November 12 - All exception types are now mapped to equivalent classes in Python. The class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception in both C++ and Python) so all IMP exceptions can be caught in Python using "except IMP.Exception". For convenience, the IMP::IOException, IMP::ValueException and IMP::IndexException classes in Python also derive from the similarly-named standard Python classes (IOError, ValueError and IndexError respectively). ## week of October 29 Update - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested). - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box(). - Added IMP::core::VolumeRestraint. - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information. - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy(). - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite. - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs. - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such. - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were. - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added. - All containers now keep track of what they looked like on the previous time step. - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier. - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications. - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints. ## Week of October 15 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement. ## Week of October 8 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems. - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this. - IMP_check() and IMP_assert() have been replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK() respectively. Hopefully this makes it clearer when to use which. IMP::set_check_level() has been revised accordingly (to use IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and IMP_not_implemented() have been renamed to IMP_FAILURE() and IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update external code, - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL - replace IMP.CHEAP with IMP.USAGE - replace IMP_assert and IMP_check as above. ## Week of October 1 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used. - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them. - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly. - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too. ## Week of September 27 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs. - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies. - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking. - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins. ## Week of September 20 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier. - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof. - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always. - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version. ## Week of August 30 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles in a box. - str should now work better in Python when applied to IMP::Object objects and a few others. Objects should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show functions directory to make it so that they work in Python. It is now a bug to \c print or \c str to not return something useful in Python for any IMP object. ## Week of August 23 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur. - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names. - Python documentation is now parsed by doxygen ## Week of August 15 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from. ## Week of August 8 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue. - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements). - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that. - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation. - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency. - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate. - The contents of bin are moving to tools since the two directories both contained assorted development tools (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon. - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored). Anyone who implemented a decorator will have to delete the second argument of their show function. ## Week of August 1 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(), IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(), IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply(). - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle() and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which updates the names in all code in subdirectories of the current directory. Please backup things first. - a function, IMP::Restraint::safe_evaluate() has been added to be used when requesting the value of restraints outside of the IMP::Model::evaluate(). ## Week of July 18 - The atom::BondPairContainer has been renamed to atom::BondPairFilter since it wasn't a real container. A real atom::BondPairContainer has been added. - Out of source builds mostly works. To do an out of source build, make a link to \c scons_tools and \c SConstruct from the source directory in the build directory and then add the line \command{repository="../path_to_src_dir"} in the \c config.py in the build directory. Then running \c scons in the build directory will put generated files there. This allows you to make both a \c release and \c fast build off of the same set of sources. Proper docs will be added soon. - The SConscript setup for a module has been greatly simplified so it should be easy enough for users to maintain their own module and updates to the build system shouldn't involve changes to user SConscripts any more. This does mean that any non-svn modules need to be updated. Look at IMP.core or another module and copy what is there. Again, proper docs will be added soon. ## Week of July 11 - Per Javi's suggestion, the conventions have been split into "Coding conventions" and "API conventions" and linked from the main page. - benchmarks got moved to their own module, IMP::benchmark and will now be tracked over time ## Week of July 4 - FilteredListSingletonContainer and FilteredPairSingletonContainer went away as there was too much hard to factor code shared with IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The IMP::core::ClosePairsScoreState is now handled by a list of filters on the class itself. That is, do IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a filter. - Hooks have been added to allow actions to be taken on an assertion or check failure. See IMP::HandleFailure and its descendants for more details. - IMP::display::CGOWriter has been much improved and now allows provides lots of display-time control and easy building of animations. - Generic optimizer states have been added (IMP::core::SingletonsOptimizerState and IMP::core::PairsOptimizerState). ## Week of June 15 - box distributions (IMP::algebra::random_vector_on_box(), IMP::algebra::random_vector_in_box()) take bounding boxes - IMP::em::DensityMap are now IMP::Object objects. Some current allocations on the stack may need to be changed (since ref counted objects can't be stack allocated). - IMP::em::DensityMap::Read() is now deprecated (as is IMP::em::DensityMap::Write) as they violated a slew of conventions. Use IMP::em::read_map() and IMP::em::write_map() instead. ## Week of June 8 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds(); - Hierarchys to IMP::atom::Hierarchies - IMP::Particles has been changed to a new type. It no longer plays quite so well with Python lists (you can't create one from an arbitrary Python list). - The containers for multiple decorators have been revised to make them more useful. See IMP::Decorators. You can now implicitly convert, for %example, a IMP::core::XYZs to an IMP::Particles (as well as use IMP::Decorators::get_particles()). All methods that take a collection of particles with a certain type now instead take a IMP::Decorators. - A method that has expectations about the particles it takes should take decorators rather than particles. That is, anything which expects IMP::core::XYZ particles, should take an IMP::core::XYZs instead of IMP::Particles. - IMP_LIST() now takes an extra argument (the name of the plural container) - In C++ IMP::Decorator objects are implicitly convertible to IMP::Particle* objects. Unfortunately, making this work in Python is problematic. - More methods take and return decorators rather than IMP::Particles. This means it is clearer what their expectations are and what they produce. ## Week of May 15 - The IMP::helper module was added and various functionality moved there. See its documentation page for a list of what is there. - IMP::algebra::SphericalCoords was renamed to IMP::algebra::SphericalVector3D to be consistent with other algebra classes. - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent with the other similar classes. - We are trying to be a bit more specific about what is expected from a IMP::atom::Hierarchy. Towards that end, there are two methods for extracting a representation from a hierarchy, IMP::atom::get_detailed_representation(), and IMP::atom::get_simplified_representation() - A method has been added to produce a simplified copy of an IMP::atom::Hierarchy. The method is still a bit crude and could benefit from other people inspecting the output and saying how it should be improved. - There is now an IMP::algebra::BoundingBoxD (and IMP::algebra::BoundingBox3D) to track bounding boxes. There are methods like IMP::algebra::bound_box(Sphere3D) which gets the bounding box of a sphere. - Each IMP::Object has its own log level. If IMP::Object::set_log_level() is called with a value other than IMP::DEFAULT, the value is used for code within and called by the object. To support this in your objects you need to add IMP_OBJECT_LOG to the start of each major method. It has been added to some of the core objects. ## Week of May 8 - IMP::atom::MolecularHierarchy has been renamed to IMP::atom::Hierarchy. Use \command{sed -i.old "s\#MolecularHierarchy\#Hierarchy\#g" my_files} - the long-deprecated Name(Decorator) was removed - The decorator suffix has been removed from decorator names. - Unit conventions have been established and documented for distance, energy, force and angle. - there is a function IMP::atom::create_protein() to create a coarse grained protein. The function is crude so far, but the idea is that we should have standard ways of generating such things so that everyone can benefit from best practices. - a class IMP::atom::BondGraph has been added so Boost.Graph algorithms can be used with bond graphs - the IMP-specific doxygen commands have been documented. - the doxygen commands unstable{name} and untested{name} have been added to doxygen to be used for marking methods and functions as likely to change or not well tested. Pass the name of the class or method as name. - bonds can now be added to molecular hierarchies automatically see IMP::atom::add_bonds() ## Week of May 1 - There is a IMP::Rotation2D - You can now extract fixed Euler ZYZ angles from a IMP::Rotation3D. This code can serve as a template for extracting other Euler angle conventions if they are needed. - IMP::algebra::rotation_between_two_vectors() changed to IMP::algebra::rotation_taking_first_to_second() to make it clear which direction the rotation goes. ## Week of April 27 - IMP::Decorator derived objects are now convertible to bool. So code like \code if (d) { blah } \endcode works. - IMP::Decorator::cast now returns a null decorator (\c DecoratorName()) rather than throwing an exception. This might make it actually useful. - New IMP::statistics module for statistical methods (e.g. clustering). ## Week of April 20 - Support for taking advantage of rigid bodies in collision detection has been added via the IMP::core::RigidClosePairsFinder. - All IMP::core::ClosePairsScoreState-related functionality has been optimized and should be noticeably faster. - The EMBED project has been merged into the IMP::em module. Thus, it is no longer necessary to obtain EMBED separately to build IMP::em. - All C++ methods which work with std::istream or std::ostream arguments should now be usable from Python: any Python file-like object can be used. - The SWIG interface now ties in to the reference counting mechanism used in the C++ code, for all classes derived from IMP::RefCounted. So, for %example, deleting a Python object should not free the underlying C++ object if another C++ object holds a reference to it, and unrefing a C++ object should not delete it if a Python object still holds a reference to it. There should be no further need to manually keep references to Python objects to prevent things from being prematurely cleaned up. This should be transparent to the user; please report any problems. - IMP::atom::AtomDecorator no longer inherits from IMP::core::XYZDecorator. ## Week of April 13 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator, IMP::atom::ResidueDecorator etc. now inherit from IMP::atom::MolecularHierarchyDecorator. This means they can set their IMP::atom::MolecularHierarchyDecorator::Type properly and saves a lot of typing in the common case. We can always make hierarchy decorators zero memory overhead if we find that the cost matters. - The base classes IMP::Comparable, IMP::NullDefault, IMP::ValidDefault and IMP::UninitializedDefault were added to document what the default constructors do for various objects in IMP. ## Week of April 6 - IMP::Object now defines pure virtual methods IMP::Object::get_version_info() and IMP::Object::show(). All non-abstract classes which inherit from IMP::Object must implement them. Doing this means they are more uniformly implemented and are documented in one place. - IMP::Object is now reference counted (the only objects this changes are IMP::Model and IMP::Optimizer). As a result optimizers and IMP::Model objects must be created on the heap. - IMP::RefCounted has had its name changed and no longer inherits from IMP::Object. This makes it easier to use independently. - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator have been improved - running doxygen produces no warnings. Let's try to keep it that way. Run \command{scons doc > /dev/null} before you check in large blocks of new code to make sure that you didn't introduce warnings. - Helper classes have been added to make SingletonModifier and PairModifier objects look like functors for C++ and Python. This makes it so you can use \c map in Python and \c for_each in C++. See IMP::SingletonFunctor and IMP::PairFunctor for more information. - The recent changes page has moved to a doxygen page to make it easier to link changes back to the relevant source code. ## Week of March 20: - create_rigid_body(), create_rigid_bodies(), create_centroid() etc. now require you to add the score state to the model yourself. Whether the rigid body is enforced or not can be controlled by adding or removing the returned IMP::ScoreState. - All previously deprecated code has been removed. This includes indexes in particles, the no-argument constructor for particles and the functionality that was moved from core to atom. - The IMP::core::RigidBodyDecorator and IMP::core::RigidMemberDecorator have been simplified by removing the traits classes. There is now only one type of rigid body and it stores the list of member particles internally. - IMP::core::NameDecorator has been deprecated as IMP::Particles have built in names (use IMP::Particle::get_name(), IMP::Particle::set_name()). - IMP::atom::ChainDecorator has been added ## Week of March 13: - A new IMP::display::Writer, IMP::display::ChimeraWriter has been added which exports geometry to native Chimera constructs. It probably should be used instead of bild and cmm for most purposes as it combines the best of both types. - compose() for transformation is now a free function so compose(rotation, vector) will work via promotions. - the geometry classes have been cleaned up and are now more consistent with IMP standards (and more likely to do what their names/comments suggest). See IMP::algebra::Sphere3D, IMP::algebra::Cylinder3D, IMP::algebra::Segment3D, IMP::algebra::Plane3D etc. - ParticleRefiner has been renamed to IMP::Refiner since the Particle was long and unnecessary. €eIF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3). - IMP::domino::DiscreteSampler::set_restraints() now allows the set of restraints used in DOMINO filtering to be controlled (previously, all restraints in the IMP::Model were used). - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the MODELLER score, not the implicit (and deprecated) score of the entire IMP::Model. - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently treated (they are supposed to be the centers of the voxels, but in some cases they were the lower left edge). The fix may result in maps and their bounding boxes appearing shifted by half the voxel size. # 2.4.0 - 2015-02-27 # {#changelog_2_4_0} - Python 3 is now supported. The Windows package now works out of the box with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu and Fedora. (Apple does not yet provide Python 3 so the Mac package is not built with Python 3 support, but if you install Python 3 you can compile IMP from source to work with it.) - A new command line tool, `rrt_sample`, allows for running rapidly exploring random tree (RRT) based sampling, e.g. to sample conformations of multiple domains connected by flexible linkers. - Command line tools are now placed in the `bin` directory of modules (the old `bin` directory, used for utility programs that are not installed, is now called `utility`). The `applications` directory is removed. - Several Python modules that were only used by EMageFit have been moved from IMP::em2d to IMP::EMageFit. - A new tutorial that uses the IMP::pmi module is now available (modeling of the RNA Polymerase II stalk). # 2.3.1 - 2014-12-01 # {#changelog_2_3_1} - Minor bug fixes. - Update to IMP::pmi to support the Nup84 and rnapolii biological systems. # 2.3.0 - 2014-10-24 # {#changelog_2_3_0} - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental. - Binary installers for 64-bit Windows are now available. - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications. - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used. - The integrative docking (idock) application now employs SOAP to improve the scoring. - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments. - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python. # 2.2.1 - 2014-08-11 # {#changelog_2_2_1} - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/) # 2.2.0 - 2014-04-09 # {#changelog_2_2_0} - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that. - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back. - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry(). - The IMP::base::Flag class was added to simplify adding flags in C++. - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers. - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones. - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls. - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application. - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead. - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added. - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available. - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples. - The container implementations has been cleaned up and simplified a bit. This should not change most usage. - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`). - [Eigen](#Eigen) is now included with IMP. - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead. - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose. - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer. - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h. - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string. - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/. # 2.1.1 - 2013-11-04 # {#changelog_2_1_1} - Added support for Mac OS 10.9 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0} - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time. - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it. - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member. - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before). - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc. - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput. - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation. - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before. - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs. - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages. - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`). - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on. - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested. - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up. - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized. # For IMP 2.0 and earlier # {#changelog_2_0} ## Spring 2013 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned. - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module. - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory. - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS. ## Winter 2012-2013 - Scons support was removed. Use cmake to build IMP, it is way better. - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long. - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody. - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols. - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo. - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems. - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake). - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit. - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file. - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice. - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript. - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case. - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed. - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info. - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far. - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin. - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code. - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page. - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python. - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated. - IMP::base::FailureHandler has been removed as it wasn't used. - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer. - Modules now must be in a folder called `modules`, which must have a `description.py` file. - The build system should now be a little better about cleaning up after itself. - How headers/doc files/data/examples - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name. - Nice version numbers will now be generated if you use git. You need to rerun tools/git/developer_tools/init_git if you have an existing clone. ## Fall 2012 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed. - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro. - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively. - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason. - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know. - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting. - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles. - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container. - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this. - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes. ## Summer 2012 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro. - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API. - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons. - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors. - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling. - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names. ## Winter 2011/2012 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy. - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()). - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes. - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores. - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input). - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use. - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented. - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed. - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects. - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct. - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole. - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed. - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer). - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions. - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified. - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated. - 24/1: \c tools/make-source has been added that adds a source/header pair to a module - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD. - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use. - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations. - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well. - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided) - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values. ## Fall 2011 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector). - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector. - IMP::statistics::HistogramD now supports display with \c matplotlib. - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply. - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn. - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom. - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel. - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d() - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support. - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility. ## Summer 2011 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy. - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone. - IMP::core::ExcludedVolumeRestraint now supports filters - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before. - you can now pick which Python to use with the \c python argument to scons. - the units for BD diffusion coefficients changed to be consistent with other IMP units. - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb. - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform. - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types. - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably. - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated. - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code. - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule. - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step. - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it. - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors. - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants. - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster. - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values. - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference(). - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially faster when used with IMP::Model::evaluate_if_good() - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value. - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder. - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent. - the RMF helper binaries have been moved to their own application so that they get installed. - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written. ## Spring 2011 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster. - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute domino on a cluster as well as debugging the sampling process. - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf. - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz. - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints. - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support. - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them. - The Python graph interface now has a remove_vertex method. - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState. - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier. - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes. ## Winter 2011 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step() - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing. - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful). - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD. - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score. - each IMP::Restraint now has its own weight. This is lightly tested. - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward. - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested). - the name of the clustering methods have been changed to create_ from get_ since they create objects. - the names of the recently added metric based clustering support has been changed to Metric rather than Distance. - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript. - the fuction IMP::display::create_restraint_geometry() was added to do what its name says. - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches. - IMP::atom::DopePairScore has been added to IMP. - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified interface to the bonded part of the CHARMM forcefield. - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons. - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide. - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and residue naming to PDB-style, so CHARMM topology files can be used unmodified. - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes. - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information. - IMP::atom::CHARMMTopology::add_missing_atoms() and IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms read from a PDB file match the CHARMM topology. - building of deprecated functionality is now false by default - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP. - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath). - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead. - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles. - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP. - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is. - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call. - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists. - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info. - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value. - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards. - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this means is that they need to be created with \c new in C++ code. It also means that one is no longer prohibited from storing them (not that there is much use in doing that). - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things. - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better. - the transform based methods on reference frame which have been deprecated for a while were removed - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source. - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version ## Fall 2010 - IMP::display::PlaneGeometry was added to display planes - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to benefit from better checking. - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing. - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer() - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState. - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD. - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory. - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that. - Logging has been improved to better report the name of methods producing the log messages - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output. - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric. - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest. - applications and biological systems now have their own documentation pages - IMP::atom::Hierarchy requires radii on its leaves - IMP::atom::read_pdb add radii to the atoms - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead. - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too. - a new example was added in the kernel to show IMP used on a biological system - IMP::display::PymolWriter now supports adding labels to things - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints. - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions. - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members. - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring. - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls ## Summer 2010 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work. - Javi added his 2D EM scoring functionality as IMP::em2D - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually. - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer). - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc. - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data). - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process - a better explanation of how to use reference counted pointers has been added to IMP::Pointer. - all simple, planar polygons are now supported for display when CGAL is present ## Week of April 1, 2010 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are - types are not checked until the lists are passed to C++ - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python) - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used. - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work. - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying. ## Week of March 13, 2010 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere. - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors). ## Week of March 6, 2010 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite. ## Week of Feb 27, 2010 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis(). - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things - renamed IMP::core::root() to IMP::core::get_root() ## Week of Feb 21, 2010 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means. - various more interpolation and endian methods had their names fixed (by adding \c get_) - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid() - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d() - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions. - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>). - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points. Some good sed strings are - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_in_box(\#get_random_vector_in(\#g" - "s\#random_vector_on_box(\#get_random_vector_on(\#g" - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g" - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g" - "s\#grid_cover\#get_grid_surface_cover\#g" - "s\#uniform_coverr\#get_uniform_surface_cover\#g" . - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD. - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments. - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis() - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point) - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base. - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance() The following global replacements are probably safe (in sed expressions) - "s\#core.distance(\#core.get_distance\#g" - "s\#algebra.distance(\#algebra.get_distance\#g" - "s\#\.transform(\#.get_transformed(\#g" Only in C++ - "s\#\.rotate(\#.get_rotated(\#g" Only in C++ - "s\#rotation_from\#get_rotation_from\#g" - "s\#rotation_in\#get_rotation_in\#g" - "s\#identity_rotation(\#get_identity_rotation_3d(" - "s\#::distance(\#::get_distance(\#g" . - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry. - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future. - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair() ## Week of Feb 14, 2010 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support. - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged. - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use. - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector. - IMP::example renamed (from examples) - The IMP_USAGE_CHECK() macro lost the (unused) third argument. ## Week of February 3, 2010 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model() - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically. - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object. - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show(). - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin. - IMP.atom.create_protein() now returns the created protein - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone(). - IMP.helper.clone() has been renamed to IMP.atom.create_clone() - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner. - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added. - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages. - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement. - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used. - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume(). ## Week of January 22, 2010 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(), IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds. - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer. - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles. - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint. - Add IMP::core::MinumumPairScore (and kin). ## Week of January 1, 2010 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed. - IMP::UsageException and IMP::InternalException are now reserved for their respective checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API exceptions (which are thrown in any build mode and should be documented and tested) and helper exceptions which are only thrown when checks are turned on (and hence should not be documented) clearer. - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves. This means that incremental evaluation of restraints in restraint sets now works and that restraints in restraint sets can be individually evaluated. Existing working usage should not have changed (so \c sed should manage updates just fine), but the internal implementation is completely new, so pay attention. - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are convertible to C++ streams and implicitly constructable from both streams and file paths. As a result, functions that take them will work when called with a string for a file path, a Python file or a C++ stream. - The IMP.examples module now documents how to use the needed aspects of scons and SWIG. - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested. - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there. - boost program options is now a checked for dependency. - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them. - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented. - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file). - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas. - a new base class IMP::Constraint has been added for constraints. Various other classes have been renamed to be consistent. ## Week of December 11 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference. - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed. - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation. - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles). ## Week of December 4 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems. - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data. - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more). ## Week of November 19 - imp.xcodeproj was moved into tools and updated accordingly - the bin directory finally went away - IMP::display::ChimeraWriter now supports the latest version of Chimera - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO. - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method. - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors) ## Week of November 12 - All exception types are now mapped to equivalent classes in Python. The class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception in both C++ and Python) so all IMP exceptions can be caught in Python using "except IMP.Exception". For convenience, the IMP::IOException, IMP::ValueException and IMP::IndexException classes in Python also derive from the similarly-named standard Python classes (IOError, ValueError and IndexError respectively). ## week of October 29 Update - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested). - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box(). - Added IMP::core::VolumeRestraint. - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information. - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy(). - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite. - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs. - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such. - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were. - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added. - All containers now keep track of what they looked like on the previous time step. - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier. - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications. - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints. ## Week of October 15 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement. ## Week of October 8 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems. - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this. - IMP_check() and IMP_assert() have been replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK() respectively. Hopefully this makes it clearer when to use which. IMP::set_check_level() has been revised accordingly (to use IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and IMP_not_implemented() have been renamed to IMP_FAILURE() and IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update external code, - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL - replace IMP.CHEAP with IMP.USAGE - replace IMP_assert and IMP_check as above. ## Week of October 1 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used. - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them. - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly. - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too. ## Week of September 27 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs. - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies. - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking. - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins. ## Week of September 20 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier. - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof. - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always. - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version. ## Week of August 30 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles in a box. - str should now work better in Python when applied to IMP::Object objects and a few others. Objects should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show functions directory to make it so that they work in Python. It is now a bug to \c print or \c str to not return something useful in Python for any IMP object. ## Week of August 23 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur. - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names. - Python documentation is now parsed by doxygen ## Week of August 15 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from. ## Week of August 8 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue. - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements). - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that. - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation. - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency. - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate. - The contents of bin are moving to tools since the two directories both contained assorted development tools (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon. - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored). Anyone who implemented a decorator will have to delete the second argument of their show function. ## Week of August 1 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(), IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(), IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply(). - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle() and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which updates the names in all code in subdirectories of the current directory. Please backup things first. - a function, IMP::Restraint::safe_evaluate() has been added to be used when requesting the value of restraints outside of the IMP::Model::evaluate(). ## Week of July 18 - The atom::BondPairContainer has been renamed to atom::BondPairFilter since it wasn't a real container. A real atom::BondPairContainer has been added. - Out of source builds mostly works. To do an out of source build, make a link to \c scons_tools and \c SConstruct from the source directory in the build directory and then add the line \command{repository="../path_to_src_dir"} in the \c config.py in the build directory. Then running \c scons in the build directory will put generated files there. This allows you to make both a \c release and \c fast build off of the same set of sources. Proper docs will be added soon. - The SConscript setup for a module has been greatly simplified so it should be easy enough for users to maintain their own module and updates to the build system shouldn't involve changes to user SConscripts any more. This does mean that any non-svn modules need to be updated. Look at IMP.core or another module and copy what is there. Again, proper docs will be added soon. ## Week of July 11 - Per Javi's suggestion, the conventions have been split into "Coding conventions" and "API conventions" and linked from the main page. - benchmarks got moved to their own module, IMP::benchmark and will now be tracked over time ## Week of July 4 - FilteredListSingletonContainer and FilteredPairSingletonContainer went away as there was too much hard to factor code shared with IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The IMP::core::ClosePairsScoreState is now handled by a list of filters on the class itself. That is, do IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a filter. - Hooks have been added to allow actions to be taken on an assertion or check failure. See IMP::HandleFailure and its descendants for more details. - IMP::display::CGOWriter has been much improved and now allows provides lots of display-time control and easy building of animations. - Generic optimizer states have been added (IMP::core::SingletonsOptimizerState and IMP::core::PairsOptimizerState). ## Week of June 15 - box distributions (IMP::algebra::random_vector_on_box(), IMP::algebra::random_vector_in_box()) take bounding boxes - IMP::em::DensityMap are now IMP::Object objects. Some current allocations on the stack may need to be changed (since ref counted objects can't be stack allocated). - IMP::em::DensityMap::Read() is now deprecated (as is IMP::em::DensityMap::Write) as they violated a slew of conventions. Use IMP::em::read_map() and IMP::em::write_map() instead. ## Week of June 8 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds(); - Hierarchys to IMP::atom::Hierarchies - IMP::Particles has been changed to a new type. It no longer plays quite so well with Python lists (you can't create one from an arbitrary Python list). - The containers for multiple decorators have been revised to make them more useful. See IMP::Decorators. You can now implicitly convert, for %example, a IMP::core::XYZs to an IMP::Particles (as well as use IMP::Decorators::get_particles()). All methods that take a collection of particles with a certain type now instead take a IMP::Decorators. - A method that has expectations about the particles it takes should take decorators rather than particles. That is, anything which expects IMP::core::XYZ particles, should take an IMP::core::XYZs instead of IMP::Particles. - IMP_LIST() now takes an extra argument (the name of the plural container) - In C++ IMP::Decorator objects are implicitly convertible to IMP::Particle* objects. Unfortunately, making this work in Python is problematic. - More methods take and return decorators rather than IMP::Particles. This means it is clearer what their expectations are and what they produce. ## Week of May 15 - The IMP::helper module was added and various functionality moved there. See its documentation page for a list of what is there. - IMP::algebra::SphericalCoords was renamed to IMP::algebra::SphericalVector3D to be consistent with other algebra classes. - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent with the other similar classes. - We are trying to be a bit more specific about what is expected from a IMP::atom::Hierarchy. Towards that end, there are two methods for extracting a representation from a hierarchy, IMP::atom::get_detailed_representation(), and IMP::atom::get_simplified_representation() - A method has been added to produce a simplified copy of an IMP::atom::Hierarchy. The method is still a bit crude and could benefit from other people inspecting the output and saying how it should be improved. - There is now an IMP::algebra::BoundingBoxD (and IMP::algebra::BoundingBox3D) to track bounding boxes. There are methods like IMP::algebra::bound_box(Sphere3D) which gets the bounding box of a sphere. - Each IMP::Object has its own log level. If IMP::Object::set_log_level() is called with a value other than IMP::DEFAULT, the value is used for code within and called by the object. To support this in your objects you need to add IMP_OBJECT_LOG to the start of each major method. It has been added to some of the core objects. ## Week of May 8 - IMP::atom::MolecularHierarchy has been renamed to IMP::atom::Hierarchy. Use \command{sed -i.old "s\#MolecularHierarchy\#Hierarchy\#g" my_files} - the long-deprecated Name(Decorator) was removed - The decorator suffix has been removed from decorator names. - Unit conventions have been established and documented for distance, energy, force and angle. - there is a function IMP::atom::create_protein() to create a coarse grained protein. The function is crude so far, but the idea is that we should have standard ways of generating such things so that everyone can benefit from best practices. - a class IMP::atom::BondGraph has been added so Boost.Graph algorithms can be used with bond graphs - the IMP-specific doxygen commands have been documented. - the doxygen commands unstable{name} and untested{name} have been added to doxygen to be used for marking methods and functions as likely to change or not well tested. Pass the name of the class or method as name. - bonds can now be added to molecular hierarchies automatically see IMP::atom::add_bonds() ## Week of May 1 - There is a IMP::Rotation2D - You can now extract fixed Euler ZYZ angles from a IMP::Rotation3D. This code can serve as a template for extracting other Euler angle conventions if they are needed. - IMP::algebra::rotation_between_two_vectors() changed to IMP::algebra::rotation_taking_first_to_second() to make it clear which direction the rotation goes. ## Week of April 27 - IMP::Decorator derived objects are now convertible to bool. So code like \code if (d) { blah } \endcode works. - IMP::Decorator::cast now returns a null decorator (\c DecoratorName()) rather than throwing an exception. This might make it actually useful. - New IMP::statistics module for statistical methods (e.g. clustering). ## Week of April 20 - Support for taking advantage of rigid bodies in collision detection has been added via the IMP::core::RigidClosePairsFinder. - All IMP::core::ClosePairsScoreState-related functionality has been optimized and should be noticeably faster. - The EMBED project has been merged into the IMP::em module. Thus, it is no longer necessary to obtain EMBED separately to build IMP::em. - All C++ methods which work with std::istream or std::ostream arguments should now be usable from Python: any Python file-like object can be used. - The SWIG interface now ties in to the reference counting mechanism used in the C++ code, for all classes derived from IMP::RefCounted. So, for %example, deleting a Python object should not free the underlying C++ object if another C++ object holds a reference to it, and unrefing a C++ object should not delete it if a Python object still holds a reference to it. There should be no further need to manually keep references to Python objects to prevent things from being prematurely cleaned up. This should be transparent to the user; please report any problems. - IMP::atom::AtomDecorator no longer inherits from IMP::core::XYZDecorator. ## Week of April 13 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator, IMP::atom::ResidueDecorator etc. now inherit from IMP::atom::MolecularHierarchyDecorator. This means they can set their IMP::atom::MolecularHierarchyDecorator::Type properly and saves a lot of typing in the common case. We can always make hierarchy decorators zero memory overhead if we find that the cost matters. - The base classes IMP::Comparable, IMP::NullDefault, IMP::ValidDefault and IMP::UninitializedDefault were added to document what the default constructors do for various objects in IMP. ## Week of April 6 - IMP::Object now defines pure virtual methods IMP::Object::get_version_info() and IMP::Object::show(). All non-abstract classes which inherit from IMP::Object must implement them. Doing this means they are more uniformly implemented and are documented in one place. - IMP::Object is now reference counted (the only objects this changes are IMP::Model and IMP::Optimizer). As a result optimizers and IMP::Model objects must be created on the heap. - IMP::RefCounted has had its name changed and no longer inherits from IMP::Object. This makes it easier to use independently. - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator have been improved - running doxygen produces no warnings. Let's try to keep it that way. Run \command{scons doc > /dev/null} before you check in large blocks of new code to make sure that you didn't introduce warnings. - Helper classes have been added to make SingletonModifier and PairModifier objects look like functors for C++ and Python. This makes it so you can use \c map in Python and \c for_each in C++. See IMP::SingletonFunctor and IMP::PairFunctor for more information. - The recent changes page has moved to a doxygen page to make it easier to link changes back to the relevant source code. ## Week of March 20: - create_rigid_body(), create_rigid_bodies(), create_centroid() etc. now require you to add the score state to the model yourself. Whether the rigid body is enforced or not can be controlled by adding or removing the returned IMP::ScoreState. - All previously deprecated code has been removed. This includes indexes in particles, the no-argument constructor for particles and the functionality that was moved from core to atom. - The IMP::core::RigidBodyDecorator and IMP::core::RigidMemberDecorator have been simplified by removing the traits classes. There is now only one type of rigid body and it stores the list of member particles internally. - IMP::core::NameDecorator has been deprecated as IMP::Particles have built in names (use IMP::Particle::get_name(), IMP::Particle::set_name()). - IMP::atom::ChainDecorator has been added ## Week of March 13: - A new IMP::display::Writer, IMP::display::ChimeraWriter has been added which exports geometry to native Chimera constructs. It probably should be used instead of bild and cmm for most purposes as it combines the best of both types. - compose() for transformation is now a free function so compose(rotation, vector) will work via promotions. - the geometry classes have been cleaned up and are now more consistent with IMP standards (and more likely to do what their names/comments suggest). See IMP::algebra::Sphere3D, IMP::algebra::Cylinder3D, IMP::algebra::Segment3D, IMP::algebra::Plane3D etc. - ParticleRefiner has been renamed to IMP::Refiner since the Particle was long and unnecessary. ˘IF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3). - IMP::domino::DiscreteSampler::set_restraints() now allows the set of restraints used in DOMINO filtering to be controlled (previously, all restraints in the IMP::Model were used). - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the MODELLER score, not the implicit (and deprecated) score of the entire IMP::Model. - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently treated (they are supposed to be the centers of the voxels, but in some cases they were the lower left edge). The fix may result in maps and their bounding boxes appearing shifted by half the voxel size. # 2.4.0 - 2015-02-27 # {#changelog_2_4_0} - Python 3 is now supported. The Windows package now works out of the box with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu and Fedora. (Apple does not yet provide Python 3 so the Mac package is not built with Python 3 support, but if you install Python 3 you can compile IMP from source to work with it.) - A new command line tool, `rrt_sample`, allows for running rapidly exploring random tree (RRT) based sampling, e.g. to sample conformations of multiple domains connected by flexible linkers. - Command line tools are now placed in the `bin` directory of modules (the old `bin` directory, used for utility programs that are not installed, is now called `utility`). The `applications` directory is removed. - Several Python modules that were only used by EMageFit have been moved from IMP::em2d to IMP::EMageFit. - A new tutorial that uses the IMP::pmi module is now available (modeling of the RNA Polymerase II stalk). # 2.3.1 - 2014-12-01 # {#changelog_2_3_1} - Minor bug fixes. - Update to IMP::pmi to support the Nup84 and rnapolii biological systems. # 2.3.0 - 2014-10-24 # {#changelog_2_3_0} - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental. - Binary installers for 64-bit Windows are now available. - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications. - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used. - The integrative docking (idock) application now employs SOAP to improve the scoring. - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments. - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python. # 2.2.1 - 2014-08-11 # {#changelog_2_2_1} - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/) # 2.2.0 - 2014-04-09 # {#changelog_2_2_0} - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that. - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back. - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry(). - The IMP::base::Flag class was added to simplify adding flags in C++. - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers. - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones. - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls. - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application. - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead. - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added. - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available. - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples. - The container implementations has been cleaned up and simplified a bit. This should not change most usage. - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`). - [Eigen](#Eigen) is now included with IMP. - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead. - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose. - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer. - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h. - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string. - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/. # 2.1.1 - 2013-11-04 # {#changelog_2_1_1} - Added support for Mac OS 10.9 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0} - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time. - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it. - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member. - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before). - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc. - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput. - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation. - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before. - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs. - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages. - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`). - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on. - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested. - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up. - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized. # For IMP 2.0 and earlier # {#changelog_2_0} ## Spring 2013 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned. - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module. - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory. - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS. ## Winter 2012-2013 - Scons support was removed. Use cmake to build IMP, it is way better. - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long. - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody. - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols. - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo. - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems. - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake). - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit. - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file. - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice. - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript. - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case. - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed. - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info. - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far. - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin. - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code. - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page. - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python. - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated. - IMP::base::FailureHandler has been removed as it wasn't used. - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer. - Modules now must be in a folder called `modules`, which must have a `description.py` file. - The build system should now be a little better about cleaning up after itself. - How headers/doc files/data/examples - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name. - Nice version numbers will now be generated if you use git. You need to rerun tools/git/developer_tools/init_git if you have an existing clone. ## Fall 2012 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed. - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro. - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively. - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason. - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know. - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting. - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles. - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container. - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this. - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes. ## Summer 2012 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro. - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API. - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons. - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors. - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling. - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names. ## Winter 2011/2012 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy. - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()). - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes. - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores. - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input). - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use. - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented. - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed. - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects. - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct. - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole. - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed. - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer). - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions. - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified. - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated. - 24/1: \c tools/make-source has been added that adds a source/header pair to a module - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD. - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use. - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations. - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well. - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided) - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values. ## Fall 2011 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector). - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector. - IMP::statistics::HistogramD now supports display with \c matplotlib. - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply. - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn. - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom. - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel. - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d() - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support. - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility. ## Summer 2011 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy. - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone. - IMP::core::ExcludedVolumeRestraint now supports filters - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before. - you can now pick which Python to use with the \c python argument to scons. - the units for BD diffusion coefficients changed to be consistent with other IMP units. - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb. - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform. - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types. - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably. - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated. - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code. - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule. - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step. - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it. - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors. - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants. - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster. - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values. - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference(). - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially faster when used with IMP::Model::evaluate_if_good() - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value. - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder. - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent. - the RMF helper binaries have been moved to their own application so that they get installed. - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written. ## Spring 2011 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster. - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute domino on a cluster as well as debugging the sampling process. - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf. - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz. - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints. - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support. - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them. - The Python graph interface now has a remove_vertex method. - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState. - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier. - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes. ## Winter 2011 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step() - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing. - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful). - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD. - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score. - each IMP::Restraint now has its own weight. This is lightly tested. - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward. - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested). - the name of the clustering methods have been changed to create_ from get_ since they create objects. - the names of the recently added metric based clustering support has been changed to Metric rather than Distance. - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript. - the fuction IMP::display::create_restraint_geometry() was added to do what its name says. - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches. - IMP::atom::DopePairScore has been added to IMP. - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified interface to the bonded part of the CHARMM forcefield. - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons. - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide. - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and residue naming to PDB-style, so CHARMM topology files can be used unmodified. - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes. - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information. - IMP::atom::CHARMMTopology::add_missing_atoms() and IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms read from a PDB file match the CHARMM topology. - building of deprecated functionality is now false by default - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP. - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath). - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead. - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles. - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP. - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is. - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call. - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists. - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info. - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value. - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards. - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this means is that they need to be created with \c new in C++ code. It also means that one is no longer prohibited from storing them (not that there is much use in doing that). - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things. - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better. - the transform based methods on reference frame which have been deprecated for a while were removed - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source. - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version ## Fall 2010 - IMP::display::PlaneGeometry was added to display planes - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to benefit from better checking. - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing. - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer() - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState. - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD. - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory. - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that. - Logging has been improved to better report the name of methods producing the log messages - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output. - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric. - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest. - applications and biological systems now have their own documentation pages - IMP::atom::Hierarchy requires radii on its leaves - IMP::atom::read_pdb add radii to the atoms - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead. - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too. - a new example was added in the kernel to show IMP used on a biological system - IMP::display::PymolWriter now supports adding labels to things - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints. - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions. - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members. - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring. - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls ## Summer 2010 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work. - Javi added his 2D EM scoring functionality as IMP::em2D - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually. - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer). - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc. - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data). - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process - a better explanation of how to use reference counted pointers has been added to IMP::Pointer. - all simple, planar polygons are now supported for display when CGAL is present ## Week of April 1, 2010 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are - types are not checked until the lists are passed to C++ - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python) - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used. - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work. - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying. ## Week of March 13, 2010 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere. - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors). ## Week of March 6, 2010 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite. ## Week of Feb 27, 2010 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis(). - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things - renamed IMP::core::root() to IMP::core::get_root() ## Week of Feb 21, 2010 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means. - various more interpolation and endian methods had their names fixed (by adding \c get_) - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid() - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d() - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions. - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>). - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points. Some good sed strings are - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g" - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g" - "s\#random_vector_in_box(\#get_random_vector_in(\#g" - "s\#random_vector_on_box(\#get_random_vector_on(\#g" - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g" - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g" - "s\#grid_cover\#get_grid_surface_cover\#g" - "s\#uniform_coverr\#get_uniform_surface_cover\#g" . - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD. - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments. - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis() - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point) - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base. - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance() The following global replacements are probably safe (in sed expressions) - "s\#core.distance(\#core.get_distance\#g" - "s\#algebra.distance(\#algebra.get_distance\#g" - "s\#\.transform(\#.get_transformed(\#g" Only in C++ - "s\#\.rotate(\#.get_rotated(\#g" Only in C++ - "s\#rotation_from\#get_rotation_from\#g" - "s\#rotation_in\#get_rotation_in\#g" - "s\#identity_rotation(\#get_identity_rotation_3d(" - "s\#::distance(\#::get_distance(\#g" . - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry. - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future. - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair() ## Week of Feb 14, 2010 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support. - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged. - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use. - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector. - IMP::example renamed (from examples) - The IMP_USAGE_CHECK() macro lost the (unused) third argument. ## Week of February 3, 2010 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model() - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically. - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object. - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show(). - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin. - IMP.atom.create_protein() now returns the created protein - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone(). - IMP.helper.clone() has been renamed to IMP.atom.create_clone() - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner. - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added. - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages. - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement. - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used. - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume(). ## Week of January 22, 2010 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(), IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds. - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer. - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles. - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint. - Add IMP::core::MinumumPairScore (and kin). ## Week of January 1, 2010 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed. - IMP::UsageException and IMP::InternalException are now reserved for their respective checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API exceptions (which are thrown in any build mode and should be documented and tested) and helper exceptions which are only thrown when checks are turned on (and hence should not be documented) clearer. - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves. This means that incremental evaluation of restraints in restraint sets now works and that restraints in restraint sets can be individually evaluated. Existing working usage should not have changed (so \c sed should manage updates just fine), but the internal implementation is completely new, so pay attention. - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are convertible to C++ streams and implicitly constructable from both streams and file paths. As a result, functions that take them will work when called with a string for a file path, a Python file or a C++ stream. - The IMP.examples module now documents how to use the needed aspects of scons and SWIG. - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested. - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there. - boost program options is now a checked for dependency. - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them. - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented. - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file). - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas. - a new base class IMP::Constraint has been added for constraints. Various other classes have been renamed to be consistent. ## Week of December 11 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference. - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed. - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation. - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles). ## Week of December 4 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems. - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data. - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more). ## Week of November 19 - imp.xcodeproj was moved into tools and updated accordingly - the bin directory finally went away - IMP::display::ChimeraWriter now supports the latest version of Chimera - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO. - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method. - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors) ## Week of November 12 - All exception types are now mapped to equivalent classes in Python. The class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception in both C++ and Python) so all IMP exceptions can be caught in Python using "except IMP.Exception". For convenience, the IMP::IOException, IMP::ValueException and IMP::IndexException classes in Python also derive from the similarly-named standard Python classes (IOError, ValueError and IndexError respectively). ## week of October 29 Update - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested). - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box(). - Added IMP::core::VolumeRestraint. - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information. - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy(). - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite. - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs. - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such. - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were. - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added. - All containers now keep track of what they looked like on the previous time step. - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier. - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications. - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints. ## Week of October 15 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement. ## Week of October 8 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems. - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this. - IMP_check() and IMP_assert() have been replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK() respectively. Hopefully this makes it clearer when to use which. IMP::set_check_level() has been revised accordingly (to use IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and IMP_not_implemented() have been renamed to IMP_FAILURE() and IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update external code, - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL - replace IMP.CHEAP with IMP.USAGE - replace IMP_assert and IMP_check as above. ## Week of October 1 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used. - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them. - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly. - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too. ## Week of September 27 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs. - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies. - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking. - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins. ## Week of September 20 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier. - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof. - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always. - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version. ## Week of August 30 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles in a box. - str should now work better in Python when applied to IMP::Object objects and a few others. Objects should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show functions directory to make it so that they work in Python. It is now a bug to \c print or \c str to not return something useful in Python for any IMP object. ## Week of August 23 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur. - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names. - Python documentation is now parsed by doxygen ## Week of August 15 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from. ## Week of August 8 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue. - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements). - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that. - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation. - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency. - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate. - The contents of bin are moving to tools since the two directories both contained assorted development tools (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon. - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored). Anyone who implemented a decorator will have to delete the second argument of their show function. ## Week of August 1 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(), IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(), IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply(). - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle() and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which updates the names in all code in subdirectories of the current directory. Please backup things first. - a function, IMP::Restraint::safe_evaluate() has been added to be used when requesting the value of restraints outside of the IMP::Model::evaluate(). ## Week of July 18 - The atom::BondPairContainer has been renamed to atom::BondPairFilter since it wasn't a real container. A real atom::BondPairContainer has been added. - Out of source builds mostly works. To do an out of source build, make a link to \c scons_tools and \c SConstruct from the source directory in the build directory and then add the line \command{repository="../path_to_src_dir"} in the \c config.py in the build directory. Then running \c scons in the build directory will put generated files there. This allows you to make both a \c release and \c fast build off of the same set of sources. Proper docs will be added soon. - The SConscript setup for a module has been greatly simplified so it should be easy enough for users to maintain their own module and updates to the build system shouldn't involve changes to user SConscripts any more. This does mean that any non-svn modules need to be updated. Look at IMP.core or another module and copy what is there. Again, proper docs will be added soon. ## Week of July 11 - Per Javi's suggestion, the conventions have been split into "Coding conventions" and "API conventions" and linked from the main page. - benchmarks got moved to their own module, IMP::benchmark and will now be tracked over time ## Week of July 4 - FilteredListSingletonContainer and FilteredPairSingletonContainer went away as there was too much hard to factor code shared with IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The IMP::core::ClosePairsScoreState is now handled by a list of filters on the class itself. That is, do IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a filter. - Hooks have been added to allow actions to be taken on an assertion or check failure. See IMP::HandleFailure and its descendants for more details. - IMP::display::CGOWriter has been much improved and now allows provides lots of display-time control and easy building of animations. - Generic optimizer states have been added (IMP::core::SingletonsOptimizerState and IMP::core::PairsOptimizerState). ## Week of June 15 - box distributions (IMP::algebra::random_vector_on_box(), IMP::algebra::random_vector_in_box()) take bounding boxes - IMP::em::DensityMap are now IMP::Object objects. Some current allocations on the stack may need to be changed (since ref counted objects can't be stack allocated). - IMP::em::DensityMap::Read() is now deprecated (as is IMP::em::DensityMap::Write) as they violated a slew of conventions. Use IMP::em::read_map() and IMP::em::write_map() instead. ## Week of June 8 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds(); - Hierarchys to IMP::atom::Hierarchies - IMP::Particles has been changed to a new type. It no longer plays quite so well with Python lists (you can't create one from an arbitrary Python list). - The containers for multiple decorators have been revised to make them more useful. See IMP::Decorators. You can now implicitly convert, for %example, a IMP::core::XYZs to an IMP::Particles (as well as use IMP::Decorators::get_particles()). All methods that take a collection of particles with a certain type now instead take a IMP::Decorators. - A method that has expectations about the particles it takes should take decorators rather than particles. That is, anything which expects IMP::core::XYZ particles, should take an IMP::core::XYZs instead of IMP::Particles. - IMP_LIST() now takes an extra argument (the name of the plural container) - In C++ IMP::Decorator objects are implicitly convertible to IMP::Particle* objects. Unfortunately, making this work in Python is problematic. - More methods take and return decorators rather than IMP::Particles. This means it is clearer what their expectations are and what they produce. ## Week of May 15 - The IMP::helper module was added and various functionality moved there. See its documentation page for a list of what is there. - IMP::algebra::SphericalCoords was renamed to IMP::algebra::SphericalVector3D to be consistent with other algebra classes. - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent with the other similar classes. - We are trying to be a bit more specific about what is expected from a IMP::atom::Hierarchy. Towards that end, there are two methods for extracting a representation from a hierarchy, IMP::atom::get_detailed_representation(), and IMP::atom::get_simplified_representation() - A method has been added to produce a simplified copy of an IMP::atom::Hierarchy. The method is still a bit crude and could benefit from other people inspecting the output and saying how it should be improved. - There is now an IMP::algebra::BoundingBoxD (and IMP::algebra::BoundingBox3D) to track bounding boxes. There are methods like IMP::algebra::bound_box(Sphere3D) which gets the bounding box of a sphere. - Each IMP::Object has its own log level. If IMP::Object::set_log_level() is called with a value other than IMP::DEFAULT, the value is used for code within and called by the object. To support this in your objects you need to add IMP_OBJECT_LOG to the start of each major method. It has been added to some of the core objects. ## Week of May 8 - IMP::atom::MolecularHierarchy has been renamed to IMP::atom::Hierarchy. Use \command{sed -i.old "s\#MolecularHierarchy\#Hierarchy\#g" my_files} - the long-deprecated Name(Decorator) was removed - The decorator suffix has been removed from decorator names. - Unit conventions have been established and documented for distance, energy, force and angle. - there is a function IMP::atom::create_protein() to create a coarse grained protein. The function is crude so far, but the idea is that we should have standard ways of generating such things so that everyone can benefit from best practices. - a class IMP::atom::BondGraph has been added so Boost.Graph algorithms can be used with bond graphs - the IMP-specific doxygen commands have been documented. - the doxygen commands unstable{name} and untested{name} have been added to doxygen to be used for marking methods and functions as likely to change or not well tested. Pass the name of the class or method as name. - bonds can now be added to molecular hierarchies automatically see IMP::atom::add_bonds() ## Week of May 1 - There is a IMP::Rotation2D - You can now extract fixed Euler ZYZ angles from a IMP::Rotation3D. This code can serve as a template for extracting other Euler angle conventions if they are needed. - IMP::algebra::rotation_between_two_vectors() changed to IMP::algebra::rotation_taking_first_to_second() to make it clear which direction the rotation goes. ## Week of April 27 - IMP::Decorator derived objects are now convertible to bool. So code like \code if (d) { blah } \endcode works. - IMP::Decorator::cast now returns a null decorator (\c DecoratorName()) rather than throwing an exception. This might make it actually useful. - New IMP::statistics module for statistical methods (e.g. clustering). ## Week of April 20 - Support for taking advantage of rigid bodies in collision detection has been added via the IMP::core::RigidClosePairsFinder. - All IMP::core::ClosePairsScoreState-related functionality has been optimized and should be noticeably faster. - The EMBED project has been merged into the IMP::em module. Thus, it is no longer necessary to obtain EMBED separately to build IMP::em. - All C++ methods which work with std::istream or std::ostream arguments should now be usable from Python: any Python file-like object can be used. - The SWIG interface now ties in to the reference counting mechanism used in the C++ code, for all classes derived from IMP::RefCounted. So, for %example, deleting a Python object should not free the underlying C++ object if another C++ object holds a reference to it, and unrefing a C++ object should not delete it if a Python object still holds a reference to it. There should be no further need to manually keep references to Python objects to prevent things from being prematurely cleaned up. This should be transparent to the user; please report any problems. - IMP::atom::AtomDecorator no longer inherits from IMP::core::XYZDecorator. ## Week of April 13 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator, IMP::atom::ResidueDecorator etc. now inherit from IMP::atom::MolecularHierarchyDecorator. This means they can set their IMP::atom::MolecularHierarchyDecorator::Type properly and saves a lot of typing in the common case. We can always make hierarchy decorators zero memory overhead if we find that the cost matters. - The base classes IMP::Comparable, IMP::NullDefault, IMP::ValidDefault and IMP::UninitializedDefault were added to document what the default constructors do for various objects in IMP. ## Week of April 6 - IMP::Object now defines pure virtual methods IMP::Object::get_version_info() and IMP::Object::show(). All non-abstract classes which inherit from IMP::Object must implement them. Doing this means they are more uniformly implemented and are documented in one place. - IMP::Object is now reference counted (the only objects this changes are IMP::Model and IMP::Optimizer). As a result optimizers and IMP::Model objects must be created on the heap. - IMP::RefCounted has had its name changed and no longer inherits from IMP::Object. This makes it easier to use independently. - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator have been improved - running doxygen produces no warnings. Let's try to keep it that way. Run \command{scons doc > /dev/null} before you check in large blocks of new code to make sure that you didn't introduce warnings. - Helper classes have been added to make SingletonModifier and PairModifier objects look like functors for C++ and Python. This makes it so you can use \c map in Python and \c for_each in C++. See IMP::SingletonFunctor and IMP::PairFunctor for more information. - The recent changes page has moved to a doxygen page to make it easier to link changes back to the relevant source code. ## Week of March 20: - create_rigid_body(), create_rigid_bodies(), create_centroid() etc. now require you to add the score state to the model yourself. Whether the rigid body is enforced or not can be controlled by adding or removing the returned IMP::ScoreState. - All previously deprecated code has been removed. This includes indexes in particles, the no-argument constructor for particles and the functionality that was moved from core to atom. - The IMP::core::RigidBodyDecorator and IMP::core::RigidMemberDecorator have been simplified by removing the traits classes. There is now only one type of rigid body and it stores the list of member particles internally. - IMP::core::NameDecorator has been deprecated as IMP::Particles have built in names (use IMP::Particle::get_name(), IMP::Particle::set_name()). - IMP::atom::ChainDecorator has been added ## Week of March 13: - A new IMP::display::Writer, IMP::display::ChimeraWriter has been added which exports geometry to native Chimera constructs. It probably should be used instead of bild and cmm for most purposes as it combines the best of both types. - compose() for transformation is now a free function so compose(rotation, vector) will work via promotions. - the geometry classes have been cleaned up and are now more consistent with IMP standards (and more likely to do what their names/comments suggest). See IMP::algebra::Sphere3D, IMP::algebra::Cylinder3D, IMP::algebra::Segment3D, IMP::algebra::Plane3D etc. - ParticleRefiner has been renamed to IMP::Refiner since the Particle was long and unnecessary. eIF3 translation initiation complex](http://salilab.org/40S-eIF1-eIF3).
255 - IMP::domino::DiscreteSampler::set_restraints() now allows the set of
256  restraints used in DOMINO filtering to be controlled (previously, all
257  restraints in the IMP::Model were used).
258 - IMP::modeller::IMPRestraints now transfers an IMP::ScoringFunction into the
259  MODELLER score, not the implicit (and deprecated) score of the
260  entire IMP::Model.
261 - Bugfix #775: coordinates of voxels in EM maps were previously inconsistently
262  treated (they are supposed to be the centers of the voxels, but in some
263  cases they were the lower left edge). The fix may result in maps and their
264  bounding boxes appearing shifted by half the voxel size.
265 
266 # 2.4.0 - 2015-02-27 # {#changelog_2_4_0}
267 - Python 3 is now supported. The Windows package now works out of the box
268  with Python 3.3 or 3.4; a `-python3` subpackage is provided for Ubuntu
269  and Fedora. (Apple does not yet provide Python 3 so the Mac package is not
270  built with Python 3 support, but if you install Python 3 you can compile
271  IMP from source to work with it.)
272 - A new command line tool, `rrt_sample`, allows for running rapidly exploring
273  random tree (RRT) based sampling, e.g. to sample conformations of multiple
274  domains connected by flexible linkers.
275 - Command line tools are now placed in the `bin` directory of modules (the
276  old `bin` directory, used for utility programs that are not installed, is
277  now called `utility`). The `applications` directory is removed.
278 - Several Python modules that were only used by EMageFit have been moved from
279  IMP::em2d to IMP::EMageFit.
280 - A new tutorial that uses the IMP::pmi module is now available (modeling of
281  the RNA Polymerase II stalk).
282 
283 # 2.3.1 - 2014-12-01 # {#changelog_2_3_1}
284 - Minor bug fixes.
285 - Update to IMP::pmi to support the Nup84 and rnapolii biological systems.
286 
287 # 2.3.0 - 2014-10-24 # {#changelog_2_3_0}
288 - A new IMP::pmi module was added, containing high-level Python classes to simplify the representation, scoring, sampling and analyzing of models in IMP. This module is still under heavy development, and should be considered experimental.
289 - Binary installers for 64-bit Windows are now available.
290 - The `.py` extension has been removed from all Python applications, so that they are consistent with the C++ applications.
291 - IMP::atom::Selection objects can now be combined using basic set operations (union, intersection, difference, symmetric difference). In Python the equivalent operators (|, &, -, ^ respectively) can be used.
292 - The integrative docking (idock) application now employs SOAP to improve the scoring.
293 - IMP::atom::CHARMMParameters::create_topology() now works with a wider range of hierarchies (previously it would only accept Residues that were direct children of Chains). For example, it should do the right thing when given a Hierarchy that contains no Chain particles, or one where Residues are grouped under Fragments.
294 - Only Python 2.6 or later is now supported. We still provide IMP for RedHat Enterprise 5 (which ships with Python 2.4) but you will need to use the `python26` binary (part of EPEL) rather than the system default Python.
295 
296 # 2.2.1 - 2014-08-11 # {#changelog_2_2_1}
297 - Bugfix: PoseScore (see the ligand_score application) should now match that used by the Pose&Rank web server (http://salilab.org/poseandrank/)
298 
299 # 2.2.0 - 2014-04-09 # {#changelog_2_2_0}
300 - A module IMP.template was added to act as a template for creating new external modules. It includes setting up git hooks for format checks and things like that.
301 - The Python support for grids was cleaned up. Hopefully this makes it more robust. If you were using a grid in Python that no longer exists, ping us, we can add it back.
302 - The generic geometry functions were simplified slightly and renamed to IMP::algebra::get_vector_geometry().
303 - The IMP::base::Flag class was added to simplify adding flags in C++.
304 - The `--log_level` and `--check_level` command line argument now take strings (eg VERBOSE) to specify their values, rather than cryptic (and a bit unstable) numbers.
305 - Command line options are now divided into normal and advanced ones. You do `--help_advanced` to show the advanced ones.
306 - Support for computing Connolly surfaces was added with IMP::algebra::get_connolly_surface(). In addition, there is now IMP::algebra::get_uniform_surface_cover(const IMP::algebra::Sphere3Ds&,double) to sample the surface of a collection of balls.
307 - IMP::atom::reate_simplified_from_volume() was added to create a simplified representation of a hierarchy that preserves the surface are and occupied volume. It uses IMP::algebra::get_simplified_from_volume(). If you want to play with such representations, you can use the rmf_simply application.
308 - IMP::atom::get_rmsd(IMP::atom::Selection, IMP::atom::Selection, const IMP::algebra::Transformation3D&) was deprecated. Use IMP::atom::get_rmsd_transforming_first() instead.
309 - IMP::algebra::get_rmsd() and IMP::algebra::get_rmsd_transforming_first() were added.
310 - The family of macros IMP_BASE_[SMALL,LARGE]_[UNORDERED,ORDERED]_[MAP,SET] have been added (eg IMP_BASE_SMALL_ORDERED_SET ()). They provide platform and boost-version independent ways to take advantage of the best set/map for an application. In particular, they use boost::container::flat_set and kin when available.
311 - IMP::atom::State and IMP::atom::Representation decorators have been added to provide support for multiple states and multiple representation schemes. They are integrated with RMFs and IMP::atom::Selection. See atom/multiresolution.py and atom/multistate.py for examples.
312 - The container implementations has been cleaned up and simplified a bit. This should not change most usage.
313 - [RMF](http://salilab.github.io/rmf) has been updated. There are a few minor API changes, in addition to a new, faster format (`.rmf3`).
314 - [Eigen](#Eigen) is now included with IMP.
315 - The deprecated function IMP::kernel::Optimizer::set_restraints() was removed. Use IMP::kernel::Optimizer::set_scoring_function() instead.
316 - The deprecated constructor for IMP::kernel::ModelObject that does not take an IMP::kernel::Model was removed. This propagates to constructors IMP::kernel::Restraint, IMP::kernel::ScoreState and kin. Make sure you pass an IMP::kernel::Model as the first constructor argument. Related to this, deprecated functions like get_is_part_of_model(), set_model() were removed as they don't serve any purpose.
317 - The deprecated methods in IMP::kernel::Optimizer that pertain to attribute optimization were removed. Inherit from IMP::kernel::AttributeOptimizer instead if you want to use them in your optimizer.
318 - Some deprecated functionality was removed. If you encounter problems, replace IMP::base::OwnerPointer with IMP::base::PointerMember, find logging stuff in IMP/base/log.h, object in IMP/base/Object.h and random in IMP/base/random.h.
319 - IMP::atom::Chain now uses strings for the chain ids, to be consistent with changes in the PDB. PDB files are written using the first character of the chain string.
320 - Added orientation-dependent SOAP scoring (IMP::score_functor::OrientedSoap and IMP::atom::OrientedSoapPairScore). Orientation-dependent SOAP scores include those that score loops (SOAP-Loop), protein-peptide interfaces (SOAP-Peptide) and proteins (SOAP-Protein). The library files themselves are rather large (~1.5GB) and so are not included with IMP. They can be downloaded separately from http://salilab.org/SOAP/.
321 
322 # 2.1.1 - 2013-11-04 # {#changelog_2_1_1}
323 - Added support for Mac OS 10.9
324 
325 # 2.1.0 - 2013-10-24 # {#changelog_2_1_0}
326 - There is now some limited IMP-wide gathering of timing statistics. You can add `--statistics=filename` to the command line or use the IMP::base::show_timings() function to view a summary of where IMP has spent time.
327 - Slack was being used inconsistently (and not always correctly) among the various containers and restraints. If you had heavily optimized the slack value for speed, you may want to revisit it.
328 - Class name OwnerPointer was switched to PointerMember, to denote more clearly a ref-counting pointer that is a class member.
329 - The dependency handling in IMP has been rewritten to simplify it and fix a bunch of bugs. There are now more checks of correctness, so behavior that might have silently worked (or not worked) before may now trigger errors. In particular, restraints/score states must have access to the IMP::kernel::Model to do much of anything, so either use constructors that take the model or call IMP::kernel::ModelObject::set_model(). Also, changing your set of inputs or outputs without calling IMP::kernel::ModelObject::set_has_dependencies(false) is now likely to trigger a check failure (it would just silently do the wrong thing before).
330 - Decorators have been cleaned up to make them behave consistently regarding IMP::kernel::Particle vs IMP::kernel::ParticleIndexes. This change shouldn't break existing code, but decorators should move to the new macros IMP_DECORATOR_METHODS(), IMP_DECORATOR_SETUP_0() etc.
331 - The IMP_GRAPH() and IMP_WEIGHTED_GRAPH() macros got an extra argument describing how to show the vertexes in the graph. They also now expose a method, eg IMP::kernel::show_as_graphviz() that writes the graph in graphviz format to an IMP::base::TextOutput.
332 - Support and rules for deprecating code in IMP have been updated. There are new macros and instructions on how to use them that should result in much better visibility about changes. See the C++ wiki for IMP for more information. And there is now an encoded policy of keeping deprecated things around for 1 release after deprecation.
333 - IMP is now much more selective about when it invalidates the dependency graph. See [Dependencies](http://github.com/salilab/imp/wiki/Dependencies) for more information. This change should make evaluation of isolated restraints much faster in certain situations where they were unexpectedly slow before.
334 - IMP::atom::Selection has been completely rewritten as it was very brittle. The set of particles returned may change in some cases, in some cases because it was buggy before, in some cases because of new bugs.
335 - Reminder: the IMP_PROTECTED macros didn't work out well, and are now deprecated. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading `_` on the method name, if you get `Swig director pure virtual method called IMP::kernel::PairScore::do_get_inputs` messages.
336 - The generated `cmake` files are no longer stored in the repository. Instead `tools/build/setup_cmake.py` is run after checkouts and other git events that change the repository tree (assuming you have run `setup_git.py`).
337 - Generation of documentation using `doxygen` has been refactored. It is now done on a per-module basis, via targets like `IMP.core-doc`. The main advantage is that changes to the doc for a module can be tested comparatively quickly, just by building that module's docs. On the down side, docs are no longer quite so heavily interconnected, so dependencies only work to things that the module actually depends on.
338 - The restraint and score state statistics functions were removed from IMP::Model. They had been broken for quite a while, with no one complaining. Statistics can be added back easily if requested.
339 - Added an IMP-specific fork of `git-flow`. It handles things like `README.md` files when you start and finish feature branches and provides a tool to nicely update IMP and display the change log. The main commands of use are `git imp feature start <feature_name>`, `git imp feature finish <feature_name>`, and `git imp update`. You may want to remove the `git-flow` lines from your `.git/config` file as you may accidentally type `git flow` instead of `git imp`. You need to run `setup_git.py` to set things up.
340 - [RMF](http://salilab.github.com/rmf) is now included as a git submodule rather than copied into the IMP repository. You should rerun `setup_git.py` to make sure submodule stuff is initialized.
341 
342 
343 # For IMP 2.0 and earlier # {#changelog_2_0}
344 ## Spring 2013
345 - Modules and applications now have a _Maintainer_ in their `README.md`, giving the github id of the person to whom issues should be assigned.
346 - `setup_git.py` supports the `--module` flag which will set up git hooks and things for a git repository that contains a module.
347 - To clean up for the release, the module dependencies are now described in a file called `dependencies.py` and the module docs in a `README.md`, both in the main module directory.
348 - IMP-wide support for progress bars was added with the IMP::base::set_progress_display() function and helpers. These are only displayed when the IMP::base::get_log_level() is IMP::base::PROGRESS.
349 
350 ## Winter 2012-2013
351 - Scons support was removed. Use cmake to build IMP, it is way better.
352 - Examples are run with a `--test` argument when IMP is built with internal checks. Expensive examples should check this flag and shorten what they do in order to try to test all their code without taking too long.
353 - The class IMP::core::NonRigidMember was added to allow particles that are not rigid but whose coordinates are specified via internal coordinates in some IMP::algebra::Reference3D that is part of an IMP::core::RigidBody.
354 - The IMP_BIG_ENDIAN and IMP_LITTLE_ENDIAN preprocessor symbols are deprecated in favor of the BOOST_BIG_ENDIAN and BOOST_LITTLE_ENDIAN symbols.
355 - Support for asymmetric moves was added to IMP::core::MonteCarlo. To do this, IMP::core::MonteCarloMover was added and IMP::core::Mover and IMP::core::MoverBase were deprecated and implemented in terms of the new class. This should require few or no changes in code that uses Monte Carlo support or implements movers (although movers should move to IMP::core::MonteCarlo). All IMP::core::MonteCarloMover objects keep track of statistics (number of moves and number of accepted moves), rather than that being done in the IMP::core::SerialMover. In addition, the poorly defined move probability was removed from IMP::core::MonteCarlo.
356 - IMP::base::WarningContext now outputs warnings immediately. This change was needed as the previous behavior resulted in occasional core dumps when the libraries for log support were unloaded before the libraries where static IMP::base::WarningContexts were declared. And there was no easy way to detect such problems.
357 - You can now (probably) build IMP with [CMake on IMP wiki](https://github.com/salilab/imp/wiki/Cmake).
358 - The maximum logging level and check level can now be controlled independently of the build mode. To, for example, enable progress level logging in a fast build, add maxlog="PROGRESS" to your config. The maximum check level for `release` has been reduced to IMP::USAGE as that speeds things up a bit.
359 - The way applications are structured and specified has been regularized and simplified. Python files are automatically copied/installed and binary executables are compiled from each found `.cpp` file linked with any `.cpp` files contained in a subdirectory called `lib`. Nothing needs to be put into the SConscript file.
360 - The Python constants like IMP.algebra.has_cgal went away as swig autogenerates ones like IMP.algebra.IMP_ALGEBRA_HAS_IMP_CGAL. While the latter are a bit uglier, the constants aren't used much and having them be auto generated is nice.
361 - There have been some major changes to the build system. The main one is that the SConscripts in the directories below the module directory aren't really used. They are being removed from the repository. Now everything really is off of file locations (eg all test_* in test is a test, each *.cpp in bin becomes a program etc). This greatly simplifies the build system, which has, historically been a disaster. You can now define dependencies by dropping files into a dependency subdirectory in your module, no need to mess with the SConscript.
362 - Building IMP now always requires specifying a `repository` argument. This was always required for out of source builds, which is what one should be doing, but is now required for in source builds to, so as to make accidentally polluting your source with an in-source build harder. Use `repository="."` in that case.
363 - The HDF5 functionality in RMF was reorganized. Specifically, the headers were moved into `RMF/HDF5`, the Python functionality into `RMF_HDF5` and the code into namespace `RMF::HDF5` and all mentions of `HDF5` or `hdf5` in the class and function names were removed.
364 - Support for IMP-wide command line flags have been added to IMP::base in the base/flags.h file. They provide a simple command line parsing interface that can be used from C++ and Python. In addition, various functionality in IMP is automatically exposed on the command line when those functions are used. For example, one can control logging, checks or profiling. See IMP/base/flags.h for (some) more info.
365 - OpenMP support in IMP is officially released. All tests pass. But there has been little specific testing so far.
366 - Access to attributes using IMP::ParticleIndex and the IMP::Model now works from Python. See IMP::Model::add_attribute() and kin.
367 - IMP::ParticleKey and IMP::ParticlesKey were renamed to IMP::ParticleIndexKey and IMP::ParticleIndexesKey respectively as that is what they store. Typedefs were provided to support old code.
368 - Indexes have been added to the docs connecting classes with all the methods that return them, examples that use them and methods that take them. And all methods that are used in examples with the examples they are used in. Please check them out and report issues and ways to improve things. You can find links on the main page.
369 - The docs for a modules/biological systems/applications should now be found in an `overview.dox` in the module `doc` folder. Modules in IMP svn have been updated. For external modules, you can just move a `overview.dox` from that was previously generated into your `doc` folder and (eventually) remove the lines from your `doc/SConscript`. This makes it easier to use the full range of doxygen markup, as well as removes the need to escape doxygen markup for Python.
370 - Examples have now been simplified into full blown doxygen examples. This involves merging the `.py` and `.readme` files (putting `\#\#` in front of each line of the `.readme`) and adding a `## \\example module/examplename.py` line at the beginning. You can use `./tools/updaters/update_examples.py modules/mymodule` to update a module that was not already updated.
371 - IMP::base::FailureHandler has been removed as it wasn't used.
372 - Handling of data and includes has been simplified. All appropriate files in the corresponding directories are linked into build/installed. This is only likely to be a problem if you had broken headers sitting around `include`. If so, just comment them out, or move them some place nicer.
373 - Modules now must be in a folder called `modules`, which must have a `description.py` file.
374 - The build system should now be a little better about cleaning up after itself.
375 - How headers/doc files/data/examples
376 - The IMP_PROTECTED macros didn't work out well, and are now deprecated. Sorry. If you had Python implementations on top of IMP base classes that implement protected virtual methods, you may need to remove a leading "_" from the method name.
377 - Nice version numbers will now be generated if you use git. You need to rerun
378 tools/git/developer_tools/init_git if you have an existing clone.
379 
380 ## Fall 2012
381 - 11/9: RMF is now an external library shipped with IMP. IMP will build it and install it automatically if you have cmake version 2.8 or higher installed.
382 - 11/9: Updated support macros for IMP::Restraint and IMP::ScoreState classes have been added (IMP_RESTRAINT_2() and IMP_SCORE_STATE_2()) that make it easier to return dependencies on things other than IMP::Particle and IMP::ScoreState objects. In addition, the dependency functions on scores and modifiers have been updated similarly and standardized. In general this should not cause backward compatibility problems unless you were overriding the IMP::ModelObject::do_get_inputs() in IMP::Restraint and using the IMP_RESTRAINT() macro. In that case update to the IMP_RESTRAINT_2() macro.
383 - 11/8: Support has been added for the C++11 keywords override (IMP_OVERRIDE) and final (IMP_FINAL). They allow you to tell the compiler to produce an error if the marked method does not override/implement a parent's method and if anything overrides a given method, respectively.
384 - 11/8: IMP::SingletonModifier::apply_indexes() and IMP::SingletonScore::evaluate_indexes() now require an upper and lower bound index (you can just pass 0 and the size of the array). We can make them option if there is a good reason.
385 - 11/8: IMP::SingletonDerivativeModifier and kin were merged into IMP::SingletonModifier and kin as the only difference, the presence of the IMP::DerivativeAccumulator (whose weight was always 1) wasn't very interested and the split added lots of complexity. If this has an impact on your code, let us know.
386 - 11/8; Evaluation with containers was greatly simplified, moving many methods on containers. We can add backwards compatibility versions if that is interesting.
387 - 11/7: Predicates, scores and modifiers written in Python need to update their methods for returning inputs and outputs. They should now have a method \c _do_get_outputs(self, m, particle_indexes) which should probably just return \c [m.get_particle(i) for in in particle_indexes] to return all the passed particles.
388 - 11/7: It is now preferred that IMP::Restraint and IMP::ScoreState descendants report dependencies directly using IMP::ModelObject::do_get_inputs() and IMP::ModelObject::do_get_outputs(). To do this, you should use the IMP_RESTRAINT_2() and IMP_CONSTRAINT_2() macros instead. Methods like IMP::Restraint::get_input_particles() are now deprecated. This change was made as the old mechanism introduced a lot of extra complication and made it so that there was no way to directly depend on something that was not a Particle or a Container.
389 - 11/3: IMP::container::InContainerPairFilter and friends now treat pairs as unordered. A flag was added to control this.
390 - 9/5: IMP::base::Object classes now have a method IMP::base::Object::clear_caches() that clears all cache data in the object. This is not universally implemented, but now provides a correct path for doing that. The IMP_LIST() type macros call that method when their contents changes.
391 ## Summer 2012
392 - 7/3: IMP_PERIODIC_OPTIMIZER_STATE() was superceded by the class IMP::core::PeriodicOptimizerState as it is rather cleaner than the macro.
393 - 7/2: in order to make various base classes in IMP more Python friendly, macros IMP_PROTECTED_METHOD(), IMP_PROTECTED_CONSTRUCTOR() were added that properly expose such methods to Python. The standards checks now complain if \c protected is used an a class in the API.
394 - 6/24: Executables from \c module/bin are now put in \b build/module_bin/module. This removes a recurring source of errors in the build script due to the origin and destination directory have the same name in scons.
395 - 6/24: Benchmarks are now put int \c module/benchmark and build to \c build/benchmark/module to remove a recurring source of errors.
396 - 6/18: The IMP::benchmark module supports the GPerfTools profilers when available. Benchmark executables (things in \c bin whose name starts with \c benchmark_) that use the IMP_BENCHMARK() macro have command line flags to turn on and cpu and memory profiling.
397 - 6/12: The naming rules for .i files that are in module \c pyext directories changed. They should now be \c IMP_modulename.name.i instead of \c IMP_modulename_name.i. This is to better support underscores in names.
398 ## Winter 2011/2012
399 - 20/4: A method, IMP::atom::read_pdb(IMP::base::TextInput, int, IMP::atom::Hierarchy), was added to load coordinates from a pdb file into an existing IMP::atom::Hierarchy.
400 - 4/4: the filter types were superceded by the predicate types. You will only see a difference if you develop a filter class (eg IMP::PairFilter -derived class) or directly call methods on filters (specifically IMP::PairFilter::get_contains()).
401 - 29/3: IMP::Particle::get_has_model() has been changed to IMP::ModelObject::get_is_part_of_model() to be consistent with other classes.
402 - 28/3: IMP::core::Mover changed slightly, adding an extra method IMP::core::Mover::get_output_particles() and requiring the IMP::Model be passed to the constructor. All movers in IMP SVN were updated without any difficulty. This allows movers to use information contained in the dependency graph amongst particles so they can, for example, weight moves based on scores.
403 - 26/3: IMP::SingletonContainerInput etc were added. Methods that take these as their arguments can be passed either containers or lists of particles (or tuples), as convenient. All methods that take IMP::SingletonContainer objects should eventually move over to taking this (changing just requires changing the argument from a pointer to an Input).
404 - 22/3: Over the last week or so, which IMP headers include which other heads have been substantially simplified. This may result in compilation errors for code outside of \c svn that depended on IMP headers including other headers. In general, you want to explicitly include the header defining each thing you use.
405 - 15/3: The vector used throughout IMP is now exposed at IMP::base::Vector and what it does it documented.
406 - 10/3: IMP::domino::RestraintCache was introduced to provide a documented centralized place to handle restraint evaluation and caching in DOMINO. In the switch various bugs involving maximum scores and IMP::RestraintSet objects were fixed.
407 - 10/3: Scoring functions in IMP are now implemented in IMP::ScoringFunction objects. All restraint evaluation is handled by them and they can cache various useful information to aid in accelerating restraint evaluation. Backward compatibility interfaces have mostly been provided, please complain if some important interface went away. This has greatly simplified the code as well as fixed various bugs in various types of restraint evaluation (particularly in the handling of maximum scores on IMP::Restraint and IMP::RestraintSet objects.
408 - 29/2: The macros IMP_NAMED_TUPLE_1(),IMP_NAMED_TUPLE_2(),IMP_NAMED_TUPLE_3() were added for easy generation of simply struct for use in passing and returning values. The generated structs support comparison, output, hashing and have constructors. Use the macros instead of a boost::tuple or a a POD struct.
409 - 16/2: Restraints now are evaluated with the correct weights when evaluated by themselves as opposed to always returning their weight as seen by the model as a whole.
410 - 15/2: The IMP::algebra grid support was rearranged and simplified. Most of the changes should be transparent. However, the IMP::algebra::grids namespace got folded into IMP::algebra, which might break some code. And various (obscure) constructors of IMP::algebra::GridD were removed. All code in svn has been changed as needed.
411 - 3/2: The IMP::domino HDF5 container has been revamped. This involves splitting it into a write version (IMP::domino::WriteHDF5AssignmentContainer) and a read version (IMP::domino::ReadHDF5AssignmentContainer).
412 - 3/2: IMP::domino::WriteAssignmentContainer and IMP::domino::ReadAssignmentContainer were added to do direct binary file writing. This is faster than the HDF5 versions.
413 - 3/2: The installation instructions have been revamped and the developer instruction, in particular, have been greatly simplified.
414 - 24/1: \c tools/show-changes has been added that shows the changes in the history log since you last updated.
415 - 24/1: \c tools/make-source has been added that adds a source/header pair to a module
416 - 24/1: principal components analysis in IMP::algebra was generalized to work in any dimension. This required renaming the class to IMP::algebra::PrincipalComponentAnalysisD.
417 - 24/1: the output of the IMP::algebra::PrincipalComponentAnalysisD::show() was changed to reflect the standards (values produce compact, single line output on show). The old show has been moved to IMP::multifit
418 - 19/1: IMP::container::ConsecutivePairContainer had IMP::container::ExclusiveConsecutivePairContainer split off of it to make the two operating mode clearer and so that there could be a clear IMP::container::ExclusiveConsecutivePairFilter to use.
419 - IMP::algebra::Grid::apply was added to provide faster access to all (or, later, part of a grid). It can be an order of magnitude or more faster than using the iterators when doing simple operations.
420 - the scons config options \c pretty and \c color were added (defaulting to True). When they are true, build commands are mostly suppressed and a much briefer description of what is being done is printed out (in color). Note that \c color mode and compilation in emacs don't get along too well.
421 - IMP::algebra::ParabolicFit was renamed to IMP::algebra::ParabolicFit2D (a backwards compatibility typedef is currently provided)
422 - IMP::algebra::LinearFit was renamed to IMP::algebra::LinearFit2D. It now supports error bars on the y-values.
423 ## Fall 2011
424 - IMP::compatibility::checked_vector was renamed to IMP::compatibility::vector and made available to IMP as IMP::vector. Typedefs should use this type (IMP::vector).
425 - Various IMP::algebra plural typedefs which had been using std::vector (since they never were updated), have now been moved to IMP::vector.
426 - IMP::statistics::HistogramD now supports display with \c matplotlib.
427 - IMP::statistics::HistogramD supercedes IMP::statistics::Histogram
428 - predicate base classes have been added: IMP::SingletonPredicate, IMP::PairPredicate etc. These take particles and return an integer. Current functionality using them includes IMP::core::PredicateSingletonScore that uses a predicate to choose which score to apply.
429 - the core generic restraint and score state support was moved to the kernel to avoid breaking various scores outside of IMP svn.
430 - the IMP::display module was refactored and moved up to be a dependency of IMP::core and IMP::atom rather than vice versa. The main observable results are that you should now use IMP::atom::HierarchyGeometry, IMP::atom::SelectionGeometry, IMP::atom::HierarchiesGeometry, IMP::core::XYZRGeometry, IMP::atom::BondGeometry etc instead of finding them in IMP::display. This change means that, display can be used internally when testing and debugging functionality in IMP::core and IMP::atom and that tests that use IMP::display don't need to build all of IMP::atom.
431 - Decomposition of restraints now results in restraints that are faster to evaluate, using the generic restraint mechanism in IMP kernel.
432 - IMP::algebra::get_unit_bounding_box_d() was changed to return a bounding box with radius 1 to be consistent with IMP::algebra::get_unit_sphere_d()
433 - A general purpose scheme to associate data with tuples of hierarchy nodes was added to the RMF format. This deprecates the special purpose bond support.
434 - the way various types of geometry is stored in RMF files was changed, breaking backwards compatibility.
435 ## Summer 2011
436 - some functionality from the kernel has been spun off into a new module IMP.base. You probably won't notice anything. The only significant changes are that functions that accept or return IMP::base::TextInput and IMP::base::TextOutput must declare them with the \c base qualifier (eg base::TextInput) so that swig is happy.
437 - the source for the kernel has been moved from \c kernel to \c modules/kernel. This is unlikely to effect anyone.
438 - IMP::core::ExcludedVolumeRestraint now supports filters
439 - IMP::domino::HDF5AssignmentTable now caches internally, which makes it several times faster than before.
440 - you can now pick which Python to use with the \c python argument to scons.
441 - the units for BD diffusion coefficients changed to be consistent with other IMP units.
442 - IMP::domino::get_interaction_graph() now takes an IMP::RestraintsTemp, like other functions
443 - The assert failure handling function was documented and renamed to IMP::handle_error() so it is easier to find when you want to use gdb.
444 - to work around various issues with ref counting pointers and compiler eccentricities, we added emulation of the C++0x \c nullptr. You should prefer that to \c NULL in IMP code that you want to run cross platform.
445 - IMP::IntsList, IMP::FloatsList and IMP::StringsList have been added to replace the various scattered solutions for passing such types.
446 - IMP::Restraint::get_decomposition() and IMP::Restraint::get_current_decomposition() changed to return a single Restraint, instead of a list. This makes it easier to propagate the weight and maximum reliably.
447 - IMP now supports usage of restraints without adding them to the IMP::Model scoring function. To do this, do IMP::Restraint::set_model(), passing the model. The restraint can then be evaluated.
448 - IMP::domino::DominoSampler does not use the IMP::domino::OptimizeRestraints restraint optimization by default. Instead, if it is asked to create the IMP::domino::RestraintScoreSubsetFilterTable, it decomposes the restraints first (via IMP::create_decomposition()). This avoids several bugs and very fragile code.
449 - the IMP::core::DiameterRestraint was ignoring the radii of the particles involved and so was giving artificially low scores. It now is on the actual diameter of this molecule.
450 - the IMP::core::MonteCarlo optimizers now support incremental evaluation. This can be a lot faster when only a few particles are moved each step.
451 - all the plural types in IMP have moved to bounds checked vectors when using debug builds with gcc. If you get a non-imp error about an invalid access, it is probably some code that walks off the end of an array. Please report it.
452 - The IMP::core::MoverBase was changed to move it off of the IMP::SingletonContainer. This resulted in the IMP::core::BallMover and IMP::core::NormalMover having new constructors. Now create them from a list of particles, a list of float keys and a radius/standard deviation. This removes some complexity and inefficiencies as well is removes some scope for errors.
453 - Temp and non-temp decorator lists have been merged in C++. This was done to simplify code since no one was using the non-temp variants. Now all decorator lists are temp (non-reference counted) variants.
454 - restraints and restraint sets are now evaluated only using "canonical" weights. That is, each restraint has one weight for the model (the total over all the restraint sets it is contained in and will always be evaluated with that weight). This allows the evaluation of simple restraints to be faster.
455 - The IMP_LIST() macros have been changed slightly. If your usage needs to take action upon addition, removal or changes to the list, use the IMP_LIST_ACTION() variant in your header. The IMP_LIST_IMPL() macro no longer takes the actions, and the last three (generally empty arguments) will have to be removed from the call to get it to work. This is needed to move towards more flexible and efficient passing of lists of values.
456 - IMP::domino uses the new IMP::Model::evaluate_if_good() support to accelerate computations
457 - IMP::core::MonteCarlo can use the IMP::Model::evaluate_if_good() support to more quickly reject
458  very bad conformations. See IMP::core::MonteCarlo::set_maximum_difference().
459 - IMP::core::ExcludedVolumeRestraint has been rewritten to be faster. It is especially
460  faster when used with IMP::Model::evaluate_if_good()
461 - a new evaluation mode IMP::Model::evaluate_if_good() and IMP::Model::evaluate_if_below() has been added. It allows restraints to shortcut score computation when the score can be shown to be worse than a maximum value.
462 - as part of moving towards more efficient use of evaluate_if_good() various deprecated methods have been removed from IMP::core::ExcludedVolumeRestraint and IMP::core::RigidClosePairsFinder.
463 - incremental evaluation support was removed from IMP. The reason for doing this was that the code was complicated, created overhead even when incremental support was not used and was not commonly used. Further, decomposing the restraints and using the dependency graph yields similar results and is more transparent.
464 - the RMF helper binaries have been moved to their own application so that they get installed.
465 - IMP::display::Writer has been revised to better support writers that write multiple frames to a single file and ones that write binary data. The noticeable changes are that the existing IMP::display::Writer instances must be created with a file name. And the IMP::display::WriteOptimizerState just takes a writer rather than a write and a file name template. IMP::display::Writer::set_frame() has been added so that you can directly control the frame being written.
466 
467 ## Spring 2011
468 - An interactive mode has been added to IMP::domino::DominoSampler. The interactive mode is useful for understanding domino sampling runs and distributing domino computations across a cluster.
469 - Functions IMP::domino::set_assignments() and IMP::domino::get_assignments() were added which support
470  writing subset assignments to data fields in hdf5 files. These can be used to helping to distribute
471  domino on a cluster as well as debugging the sampling process.
472 - a Python function IMP.show_graphviz() has been added that attempts to render a graph with \c dot and then open the pdf.
473 - The Python functions IMP.show_altgraph() and IMP.get_networkx_graph() have been added to provide alternatives for quick viewing of IMP graphs. Unfortunately, they still pale compared to graphviz.
474 - IMP.atom.create_cover() was added to facilitate creating a summary particle for part of a hierarchy in order to accelerate restraints.
475 - IMP.hdf5 was renamed to IMP.rmf, the proposed name of the file format that it supports. That also better separates the hdf5 support from the file format support.
476 - The old I/O functions have been marked as deprecated in favor of IMP.hdf5. You may need to compile with deprecated="yes" if you use them.
477 - The Python graph interface now has a remove_vertex method.
478 - One can now access the model restraint evaluation statistics programmatically using IMP::Model::get_restaint_statistics(). Restraint statistics can be logged to a file during optimization using IMP::core::WriteRestraintScoreStatisticsOptimizerState.
479 - IMP::domino::DominoSampler now allows you to specify the tree on which merges occur directly. Also, you now need to specify the merge tree rather than the JunctionTree if you want to do non-loopy inference and specify something about how it is done. Sorry, it makes the code a lot easier.
480 - Preliminary symmetry support has been added via the IMP::core::TransformationSymmetry and IMP::core::TransformationAndReflectionSymmetry classes.
481 
482 ## Winter 2011
483 - There now is a common base class IMP::atom::Simulation for IMP::atom::BrownianDynamics and IMP::atom::MolecularDynamics. In merging them there were several minor changes made to both. The most noticeable may be that the function for setting the maximum time step length is now IMP::atom::Simulation::set_maximum_time_step(). The last time step length can be extracted by IMP::atom::Simulation::get_last_time_step()
484 - IMP::core::MonteCarlo::set_temperature() was renamed to IMP::core::MonteCarlo::set_kt() as that is what it is doing.
485 - A new module IMP.hdf5 has been added to provide more general support for saving and loading IMP::atom::Hierarchy structures to files. Multiple conformations of each hierarchy can be saved as well as information about geometry and restraints (although these can't be read at the moment, so they are not so useful).
486 - IMP::algebra::grids::GridD now supports dense grids in arbitrary dimension through use of IMP::algebra::grids::DenseGridStorageD.
487 - IMP::Optimizer now can use a list of restraint sets to compute the score rather than the complete Model score.
488 - each IMP::Restraint now has its own weight. This is lightly tested.
489 - A new log level IMP::PROGRESS has been added. The idea is that IMP::Sampler::get_sample() calls and other expensive loops can show a boost progress meter at this log level to give some idea when things are moving forward.
490 - IMP::domino::TrivialParticleStates renamed to IMP::domino::IndexParticleStates to be less judgmental
491 - IMP::domino::CompoundParticleStates and IMP::domino::RecursiveParticleStates added
492 - IMP::domino::DominoSampler now supports maximum scores on IMP::RestraintSets (untested).
493 - the name of the clustering methods have been changed to create_ from get_ since they create objects.
494 - the names of the recently added metric based clustering support has been changed to Metric rather than Distance.
495 - IMP now has the ability to have certain tests designated as expensive and to skip those tests when test-fast is run. The motivation for this is that the tests take a really long time making it impractical to run before every commit. And it is hard to tell which tests have not been run when running with the necessary multiple build processes. My thought is that tests that take about a second or less on a debug build should be left along (this is probably >95% of the tests), but the few that take longer should be marked as expensive. To do that, rename the test to expensive_text_XXX.py (from test_XXX.py) and add expensive_python_tests= env.IMPModuleGetExpensivePythonTests() to the IMPModuleTest call in the test SConscript.
496 - the fuction IMP::display::create_restraint_geometry() was added to do what its name says.
497 - runtime dimensional geometric objects have been added. These include IMP::algebra::VectorKD, IMP::algebra::BoundingBoxKD and IMP::algebra::SphereKD. The IMP::statistics::Embedding classes now use IMP::algebra::VectorKD instead of Floats. The sparse grids also support variable dimensions as to the nearest neighbor searches.
498 - IMP::atom::DopePairScore has been added to IMP.
499 - IMP::atom::CHARMMStereochemistryRestraint provides a high-level simplified
500 interface to the bonded part of the CHARMM forcefield.
501 - IMP modules are now built by first combining all the .cpp files into one, and then building that file. This greatly accelerates building of a module from scratch as build time is dominated by parsing headers. This does change the semantics of .cpp files as they see headers included by and functions declared by an arbitrary subset of other .cpp files. For now they are still required to be able to be compiled separately. To get the only behavior back either globally or on a per-module basis, see the percppcompilation parameter to scons.
502 - Instructions and scripts have been added for building modules and things externally to the IMP build. See the installation guide.
503 - IMP::atom::CHARMMParameters can now automatically map CHARMM-style atom and
504 residue naming to PDB-style, so CHARMM topology files can be used unmodified.
505 - IMP::atom::CHARMMTopology::add_sequence() provides a simple way to generate a topology from a string of amino acid one-letter codes.
506 - IMP::atom::CHARMMTopology::add_coordinates() will generate atomic Cartesian coordinates from CHARMM internal coordinate information.
507 - IMP::atom::CHARMMTopology::add_missing_atoms() and
508 IMP::atom::remove_charmm_untyped_atoms() can now be used to ensure that atoms
509 read from a PDB file match the CHARMM topology.
510 - building of deprecated functionality is now false by default
511 - the IMP_COMPARISONS(), IMP_COMPARISONS_1(), IMP_COMPARISONS_2() macros now take the name of the class as an additional (first) argument. This avoids the formerly needed, \c This typedef and makes them more consistent with the other macros in IMP.
512 - modules can now be built externally to the IMP svn in addition to applications and biological systems. To do so, create a directory with your module in a subdir containing links to the IMP SConscript file and scons_tools and a config.py that has information needed to run IMP (eg an appropriate \c pythonpath, \c ldpath).
513 - removed IMP::algebra::GridD::get_index(VectorD) as it wasn't very safe (in that vectors coordinates can get rounded between the IMP::algebra::GridD::get_has_index() and IMP::algebra::GridD::get_index() calls. Use IMP::algebra::GridD::get_nearest_index() or IMP::algebra::GridD::get_extended_index() or IMP::algebra::GridD::operator[]() instead.
514 - renamed IMP::domino::ParticleStatesTable::get_particles() to IMP::domino::ParticleStatesTable::get_subset() as it is more useful to be able to get the subset containing all the particles.
515 - rename IMP::restrainer::Restraint::print_all_restraints() to IMP::restrainer::Restraint::show_all_restraints() to be consistent with the rest of IMP.
516 - renamed IMP::atom::get_conect_record_string() to IMP::atom::get_pdb_conect_record_string() to make clearer what it is.
517 - class and method names are now spell checked. Add missing words to spelling_exceptions in the module test call.
518 - the plural names (eg IMP.Particles()) have been removed from the Python side as they are just Python lists.
519 - biological systems can now include internal modules. These provide a convenient place to put application specific C++ and Python code. See the example system for more info.
520 - IMP_VALUES() now includes IMP_OUTPUT_OPERATOR() so you no longer need two namespace macros for each value.
521 - the function IMP::core::Harmonic::k_from_standard_deviation was renamed to IMP::core::Harmonic::get_k_from_standard_deviation() to conform to the standards.
522 - the IMP::atom::PDBSelector and IMP::atom::Mol2Selector became IMP::Object classes. All this
523  means is that they need to be created with \c new in C++ code. It also means that one is
524  no longer prohibited from storing them (not that there is much use in doing that).
525 - the standards tests have been improved to test more useful aspects. Now class and function names are checked some as well as all modules have the checks for show and plural types. The error messages should explain how to fix things.
526 - IMP::core::RigidBody::lazy_set_reference_frame() was renamed to IMP::core::RigidBody::set_reference_frame_lazy() to fit with the standards better.
527 - the transform based methods on reference frame which have been deprecated for a while were removed
528 - the hierarchy visiting and searching functions have been renamed to better fit the IMP naming scheme as we don't expect they were used outside of the IMP source.
529 - IMP::atom::NonhydrogenMol2Selector was renamed IMP::atom::NonHydrogenMol2Selector to be consistent with the PDB version
530 
531 ## Fall 2010
532 - IMP::display::PlaneGeometry was added to display planes
533 - the optimize methods on optimizers was moved to a "non-virtual interface". Usage has not changed, but optimizers should implement IMP::Optimizer::do_optimize() instead of IMP::Optimizer::optimize() to
534 benefit from better checking.
535 - the two-vector constructor for IMP::algebra::Plane3D now expects the second vector to be a unit vector as otherwise one could switch the order around without noticing.
536 - the io functions in the kernel have been cleaned up a bit (the old ones are there for backwards compatibility). See IMP::write_particles(), IMP::write_particles_binary(), IMP::write_particles_to_buffer() and IMP::read_particles(), IMP::read_particles_binary(), IMP::read_particles_from_buffer()
537 - IMP::core::WriteModelOptimizerState is going away as it duplicated IMP::WriteParticlesOptimizerState.
538 - the IMP::algebra::Grid3D has been changed to make it support any (fixed) dimension when used with sparse storage. The template arguments have also been changed to remove a redundant one. See IMP::algebra::GridD.
539 - the biological systems specification has been improved a little adding ability to run some rudimentary tests to make sure the scripts are ok and to not require that everything be run from the svn directory.
540 - A new method IMP::atom::create_compatible_rigid_body() which creates a rigid body from a hierarchy which has the same internal coordinates as another rigid body. A way to make the rigid body construction more canonical would be nice (so we don't need this function), but I'm not sure how to do that.
541 - Logging has been improved to better report the name of methods producing the log messages
542 - a macro IMP_FUNCTION_LOG has been added which should be added to the start of non-member functions which produce significant logging output.
543 - IMP::algebra::get_distance(const Rotation3D&,const Rotation3D&) was changed to use the Euclidean distance in R4 as that seems better behaved than the prior metric.
544 - IMP.domino2 was renamed to IMP.domino and the old IMP.domino was moved
545 - the documentation for classes and some methods now includes a list of all examples which use the class. Please report any incorrect links or phantom classes appearing in the docs. Unfortunately, functions have to be added individually. Please request any functions of particular interest.
546 - applications and biological systems now have their own documentation pages
547 - IMP::atom::Hierarchy requires radii on its leaves
548 - IMP::atom::read_pdb add radii to the atoms
549 - IMP::helper was merged into IMP::restrainer as all the functionality there was primarily to provide support for restrainer
550 - IMP::domino2::SubsetEvaluatorTable was removed. Pass the model and IMP::domino2::ParticleStatesTable to the IMP::domino2::RestraintScoreSubsetFilterTable instead.
551 - IMP::atom::Selection was added to allow specification of parts of IMP::atom::Hierarchy using biological sorts of names. A few functions to create restraints from them were added too.
552 - a new example was added in the kernel to show IMP used on a biological system
553 - IMP::display::PymolWriter now supports adding labels to things
554 - IMP::display::PairRestraintGeometry and IMP::display::ConnectivityRestraintGeometry were added to aid in visualizing those restraints.
555 - IMP::bullet was added with an optimizer to use hard body physics to resolve local collisions.
556 - IMP::atom::setup_as_rigid_body() has been deprecated and replaced by IMP::atom::create_rigid_body(). The latter creates the rigid body as a separate particle so that the molecular hierarchy doesn't have to worry about having its bounding sphere large enough to enclose all the members.
557 - IMP::core::KClosePairsPairScore was added to have count based cutoffs for close pair scoring.
558 - IMP::algebra::get_surface_area_and_volume() was added to compute the two numbers for a set of balls
559 
560 ## Summer 2010
561 - The IMP::core::RigidDistancePairScore and IMP::core::ClosePairsPairScore have been revised to remove certain inconsistencies and ambiguities and to make them faster. Existing code using them will need to be slightly tweaked to work.
562 - Javi added his 2D EM scoring functionality as IMP::em2D
563 - some new ways to speed up scons have been added. See IMP::core::HarmonicDistancePairScore, IMP::core::SoftSpherePairScore and, for C++ users, the generic restraint support IMP::container::create_restraint() and IMP::core::create_restraint(). These each offer between 2x and 1.2x speedup when using simple scores (like harmonics on distances). In addition, a function IMP::container::get_slack_estimate() has been added which will estimate the correct slack to use with a IMP::container::ClosePairContainer. This will be extended to support IMP::core::ExcludedVolumeRestraint eventually.
564 - support for dense and sparse grids have been added to IMP::algebra. See the IMP::algebra module page (I'm still working on how to make the docs clearer).
565 - all write functions now have associated IMP::OptimizerState and IMP::FailureHandler classes. For examples, see IMP::WriteOptimizerState, IMP::WriteBinaryOptimizerState, IMP::WriteFailureHandler, IMP::atom::WritePDBOptimizerState etc.
566 - functions IMP::write_binary_model(), IMP::read_binary_model() and IMP::read_configurations() have been added to support space and filesystem efficient storage of model configurations. Multiple configuration can be written to a single file. This functionality requires NetCDF be installed (NetCDF is a standard library for reading and writing binary data).
567 - domino2 has been added. It provides a simpler interface on Keren's domino optimize as well as a number of new optimization and hooks for customizing the optimization process
568 - a better explanation of how to use reference counted pointers has been added to IMP::Pointer.
569 - all simple, planar polygons are now supported for display when CGAL is present
570 
571 ## Week of April 1, 2010
572 
573 - All plural objects in Python are now simply Python lists rather than C++ recreations of Python lists. This makes compilation of the wrappers much, much, much faster. The visible changes are
574  - types are not checked until the lists are passed to C++
575  - passing std::vector types by pointer and storing passed std::vector types by pointer will not work (but should never have been done anyway from code exposed to Python)
576  - swig has issues with some overloads (not always for obvious reasons). These are being sorted out. None are commonly used.
577  - get_particles() on a plural object doesn't work in Python (and is not needed in C++). Implicit conversions will always work.
578 
579 - All containers need to know the model at time of construction. If, after updating, you get an error on a container constructor, just add the model as the first argument. This also means that empty lists of particles in the constructors cause check failures. We can figure out a work around if this is too annoying.
580 
581 ## Week of March 13, 2010
582 - function get_rotation_in_radians_about_axis() was renamed to get_rotation_about_axis() as it didn't parse and we don't have the "in radians" elsewhere.
583 
584 - Added IMP::algebra::ReferenceFrame3D. As part of the cleanup, IMP::core::RigidBody::set_reference_frame() replaces set_transformation(). And the reference-frame related transformation construction functions were replaced. IMP::algebra::get_rotation_from_x_y_axes() was added to aid in the construction of reference frames (and replaces the algebra function taking three vectors).
585 
586 ## Week of March 6, 2010
587 - Publications now have a description field to give browsers an indication of what is in the publication to aid selection of what to read/cite.
588 
589 
590 ## Week of Feb 27, 2010
591 - IMP::algebra::decompose_rotation_into_axis_angle() was renamed to IMP::algebra::get_angle_and_axis().
592 
593 - renamed IMP::atom::bond() and IMP::atom::unbond() to IMP::atom::create_bond() and IMP::atom::destroy_bond() to be more consistent with things
594 
595 - renamed IMP::core::root() to IMP::core::get_root()
596 
597 
598 ## Week of Feb 21, 2010
599 
600 - IMP::Object::set_was_owned() was changed to IMP::Object::set_was_used() as that better reflects what the flag means.
601 - various more interpolation and endian methods had their names fixed (by adding \c get_)
602 - the function get_mean() on vectors was removed as it duplicated IMP::algebra::get_centroid()
603 - build_Transformation3D_from_transformation3D was replaced by IMP.algebra.get_transformation_3d()
604 - IMP::algebra::almost_equal() is now IMP::algebra::get_are_almost_equal() to fit the naming conventions.
605 - the random vector generators have been cleaned up to remove redundant ones and make the names more consistent. They now have the form of either
606  - IMP::algebra::get_random_vector_in() or IMP::algebra::get_random_vector_on() where there is one argument describing the volume or shape (such as an IMP::algebra::SphereD<4>).
607  - or IMP::algebra::get_uniform_surface_cover() where the first argument is the primitive whose surface we are covering and the second is the number of points.
608  Some good sed strings are
609  - "s\#random_vector_in_unit_sphere()\#get_random_vector_in(IMP.algebra.get_unit_sphere_3d())\#g"
610  - "s\#random_vector_in_unit_box()\#get_random_vector_in(IMP.algebra.get_unit_bounding_box_3d())\#g"
611  - "s\#random_vector_on_unit_sphere()\#get_random_vector_on(IMP.algebra.get_unit_sphere_3d())\#g"
612  - "s\#random_vector_on_unit_box()\#get_random_vector_on(IMP.algebra.get_unit_bounding_box_3d())\#g"
613  - "s\#random_vector_in_box(\#get_random_vector_in(\#g"
614  - "s\#random_vector_on_box(\#get_random_vector_on(\#g"
615  - "s\#random_vector_in_sphere(\#get_random_vector_in(\#g"
616  - "s\#random_vector_on_sphere(\#get_random_vector_on(\#g"
617  - "s\#grid_cover\#get_grid_surface_cover\#g"
618  - "s\#uniform_coverr\#get_uniform_surface_cover\#g"
619  .
620 - Spheres in IMP::algebra are now supported in general dimensions. See IMP::algebra::SphereD.
621 - Many methods in IMP::algebra were renamed to be more consistent with other methods or general conventions. Most of these simply involved adding \c get_ as a prefix and/or add a \c _3d suffix when they did not take any arguments.
622  - build_Rotation2D_from_Vector2D became IMP::algebra::get_rotation_to_x_axis()
623  - most of the Rotation3D and Transformation3D functions (in addition, IMP::algebra::get_identity_rotation_3d() and IMP::algebra::get_identity_transformation_3d() added the \c _3d so we can add 2D versions at some point)
624  - IMP::algebra::Transformation3D::get_transformed() and IMP::algebra::Rotation3D::get_rotated(). Replacing \c ".transform(" worked in the IMP code base.
625  - IMP::algebra::get_distance(), this also carries over to the IMP::core::get_distance()
626  The following global replacements are probably safe (in sed expressions)
627  - "s\#core.distance(\#core.get_distance\#g"
628  - "s\#algebra.distance(\#algebra.get_distance\#g"
629  - "s\#\.transform(\#.get_transformed(\#g" Only in C++
630  - "s\#\.rotate(\#.get_rotated(\#g" Only in C++
631  - "s\#rotation_from\#get_rotation_from\#g"
632  - "s\#rotation_in\#get_rotation_in\#g"
633  - "s\#identity_rotation(\#get_identity_rotation_3d("
634  - "s\#::distance(\#::get_distance(\#g"
635  .
636 - IMP::algebra::Vector3D and IMP::algebra::Vector3Ds were replaced by the expanded (non-typedef) version in functions parsed by SWIG as the current situation was extremely brittle. For the moment, you should not use the typedefs in definitions that swig sees. Sorry.
637 - IMP::algebra::Sphere3DPatch was renamed to IMP::algebra::SpherePatch3D to be consistent with the other geometry and allow us to support higher dimensions in the future.
638 - rigid_align_first_to_second was renamed to IMP::algebra::get_transformation_aligning_first_to_second() and build_Transformation2D_from_point_sets was renamed to IMP::algebra::get_transformation_aligning_pair()
639 
640 ## Week of Feb 14, 2010
641 
642 - The IMP::atom::Bond types were rearranged to reconcile the protein-centric classification with ligand support.
643 
644 - IMP::atom::Atom no longer has charge since we have IMP::atom::Charged.
645 
646 - IMP::atom::read_mol2() and IMP::atom::write_mol2() were added
647 
648 - IMP::atom::Selector and descendants renamed to IMP::atom::PDBSelector since we will soon have Mol2Selectors. In addition, the IMP::atom::IgnoreAlteriativesSelector is now IMP::atom::NonAlternativeSelector to keep the name positive, like the others. In the IMP body, all instances of the string Selector were for PDBSelectors and so sed was safe to use.
649 
650 - IMP::atom::read_pdb() lost the "ignore_alternatives" argument. Instead, all of the provided selectors, other than IMP::atom::AllSelector ignore alternatives internally. Code that uses the IMP::atom::AllSelector should probably change to the IMP::atom::NonAlternativeSelector.
651 
652 - IMP::example renamed (from examples)
653 
654 - The IMP_USAGE_CHECK() macro lost the (unused) third argument.
655 
656 ## Week of February 3, 2010
657 
658 - The model io code as moved into the kernel so it is now IMP::write_model() and IMP::read_model()
659 
660 - Remove the version_info argument from the implementation macros. Instead, just pick up the module version info automatically.
661 
662 - Add a function Object::get_type_name() and Object::get_module() to get the respective bits of information about each object.
663 
664 - Change Object::show() to be a non-virtual wrapper which prints the object name and other info before passing control off to Object::do_show().
665 
666 - The containers were moved to IMP.container. See that module page for a full list of what moved. In the process, the backwards compatibility functions were removed from IMP::container::PairsRestraint and kin.
667 
668 - IMP.atom.create_protein() now returns the created protein
669 
670 - the simplify functions have been moved from IMP.helper. They are now IMP.atom.create_simplified_along_backbone().
671 
672 - IMP.helper.clone() has been renamed to IMP.atom.create_clone()
673 
674 - IMP.misc.ConnectingPairContainer has been moved to IMP.core.ConnectingPairContainer
675 
676 - The locations for pixels in IMP::em::DensityMap have been moved to the center of the corresponding voxel from the lower left corner.
677 
678 - IMP::em::get_transformed(), IMP::em::get_resampled() and IMP::em::get_density() have been added.
679 
680 - The documentation has been majorly cleaned up and simplified. Please look around and see if you like it. There are now two main pages, the introduction and the developers guide, rather than the previous plethora of pages.
681 
682 - IMP::algebra::Cone3D::get_bounding_sphere() has been removed since it was not of general use, not returning the bounding sphere you would expect and was trivial to implement.
683 
684 - IMP::algebra::projection() was removed since it didn't actually project onto the segment and the name didn't follow the normal conventions. And it wasn't used.
685 
686 - Most geometry now has namespace methods IMP::algebra::get_bounding_box(), IMP::algebra::get_surface_area(), IMP::algebra::get_volume().
687 
688 ## Week of January 22, 2010
689 
690 - A generic geometry layer has been added (see \ref genericgeometry). Using it, IMP::atom::rmsd(),
691  IMP::algebra::alignment_taking_first_to_second() and the new IMP::atom::native_overlap() all
692  can work on either collections of either IMP::core::XYZ particles or IMP::algebra::Vector3Ds.
693 
694 - Deprecated functionality has been removed from IMP.core. For those still using it, IMP::core::ClosePairsScoreState was replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairsScoreState was replaced by IMP::core::CloseBipartitePairContainer.
695 
696 - The macros for Scores changed slightly. Scores not in IMP svn which use the IMP_*_SCORE macros now have to implement the get_is_changed() function outside of the class declaration and scores which use the IMP_SIMPLE_*_SCORE now don't have to implement the functions to get interactions and used particles.
697 
698 - Rename IMP::core::MinimumPairScoreRestraint (and kin, Maximum, Singleton, Triplet and Quad) to IMP::core::MinimumPairRestraint.
699 
700 - Add IMP::core::MinumumPairScore (and kin).
701 
702 ## Week of January 1, 2010
703 
704 - IMP::em is now using Particles to resample density maps and for fitting restraints. The previous ParticleAccessPoint mechanism is removed.
705 
706 - IMP::UsageException and IMP::InternalException are now reserved for their respective
707  checks (and cannot be thrown by IMP_THROW()). This makes the distinction between API
708  exceptions (which are thrown in any build mode and should be documented and tested)
709  and helper exceptions which are only thrown when checks are turned on (and hence
710  should not be documented) clearer.
711 
712 - IMP::core::RestraintSet moved to IMP::RestraintSet. The restraint sets now simply define a tree
713  with weights which the model traverses as opposed to being an opaque IMP::Restraint themselves.
714  This means that incremental evaluation of restraints in restraint sets now works and that
715  restraints in restraint sets can be individually evaluated. Existing working usage should not
716  have changed (so \c sed should manage updates just fine), but the internal
717  implementation is completely new, so pay attention.
718 
719 - IMP now has a text i/o layer consisting of IMP::TextInput and IMP::TextOutput. These classes are
720  convertible to C++ streams and implicitly constructable from both streams and file paths. As a result,
721  functions that take them will work when called with a string for a file path, a Python file or a
722  C++ stream.
723 
724 - The IMP.examples module now documents how to use the needed aspects of scons and SWIG.
725 
726 - The IMP.examples module is used as the basis for new modules created by the make-module script, so those include more documentation and are tested.
727 
728 - Module overview pages are now generated from the information passed in the SConscript file, rather than from the (now removed) modulename.dox file. See the examples file for an example of the things you can do there.
729 
730 - boost program options is now a checked for dependency.
731 
732 - The IMP::Model now uses dependency information from IMP::ScoreStates to automatically reorder the IMP::ScoreStates so that they are evaluated in the right order. Also, IMP::Model::evaluate() on a subset of the restraints now only updates the needed score states, not all of them.
733 
734 - Functions on rigid bodies no longer provide the IMP::core::RigidMemberRefiner as a default as it is, pretty much always, the wrong thing to use. That class may go away without notice in the future and is no longer documented.
735 
736 - the IMP::helper::create_simple_connectivity_on_rigid_bodies() and IMP::helper::create_simple_excluded_volume_on_rigid_bodies() now both take an IMP::Refiner as an (optional) final argument. This makes it so they can be used with the rigid bodies created from molecular hierarchies. This code still needs to be updated and the docs don't reflect what the code was or is doing, so use with caution (see notes in the header file).
737 
738 - The IMP::helper create* functions now all take non-Object inputs as const& to make clear what is input and that there are no memory management gotchas.
739 
740 - a new base class IMP::Constraint has been added for constraints. Various
741 other classes have been renamed to be consistent.
742 
743 ## Week of December 11
744 
745 - The interface of IMP::PairScore changed slightly to make the function signatures more uniform across various methods. Now all methods that take a pair of IMP::Particle objects take a const IMP::ParticlePair&. Previously some took that, some took an IMP::ParticlePair and some took two IMP::Particle*s. If you have an implementation of an IMP::PairScore outside of svn, you will have to update your function signatures (the compiler will tell you where). Backward compatible IMP::PairScore::evaluate() methods have been provided, so simple users of pair scores shouldn't see a difference.
746 
747 - IMP::core::PairRestraint takes a IMP::ParticlePair in the constructor instead of two IMP::Particle*s. A backwards compatible constructor can be added if needed.
748 
749 - IMP::ParticlePair::first and second went away since no one claimed to be using it and it complicated the implementation.
750 
751 - The generic restraints/score states/scores infrastructure has been extended to triplets and quads (for angles and torsion angles).
752 
753 ## Week of December 4
754 
755 - IMP now makes use of Boost.Filesystem to manipulate paths. Things may not work well on systems where boost libs are not available which does not support posix paths. But libless boost installs generally don't exist outside of our test systems.
756 
757 - new functions like IMP::core::get_example_path(filename) and IMP::core::get_data_path(filename) allow one to easily find a data file named 'filename' for the examples and library internal data.
758 
759 - The IMP::display module has been revamped in order to support non-linear geometry. VRML support has been dropped. CMM support will be dropped unless someone requests it (as IMP::display::ChimeraWriter does all it does and more).
760 
761 ## Week of November 19
762 
763 - imp.xcodeproj was moved into tools and updated accordingly
764 
765 - the bin directory finally went away
766 
767 - IMP::display::ChimeraWriter now supports the latest version of Chimera
768 
769 - IMP::display::CGOWriter was renamed to IMP::display::PymolWriter as it can do more than just CGO.
770 
771 - IMP::VectorOfRefCounted::operator[] can now be used for assignments just like a normal vector. No more need to use the IMP::VectorOfRefCounted::set() method.
772 
773 - the IMP::atom::Hierarchy constructor from an IMP::Particle* is now explicit (as were the decorator constructors)
774 
775 
776 ## Week of November 12
777 
778 - All exception types are now mapped to equivalent classes in Python. The
779 class hierarchy is similar (e.g. IMP::ModelException derives from IMP::Exception
780 in both C++ and Python) so all IMP exceptions can be caught in Python using
781 "except IMP.Exception". For convenience, the IMP::IOException,
782 IMP::ValueException and IMP::IndexException classes in Python also derive from
783 the similarly-named standard Python classes (IOError, ValueError and IndexError
784 respectively).
785 
786 ## week of October 29 Update
787 
788 - You can now add members to rigid bodies via IMP::core::RigidBody::add_member(). These members can be other rigid bodies (although this functionality is only lightly tested).
789 
790 - The code to get bounding volumes for IMP::atom::Hierarchy has been moved to the atom module, as as the clone and destroy code. They can be found at IMP::atom::clone(), IMP::atom::destroy(), IMP::atom::get_bounding_box().
791 
792 - Added IMP::core::VolumeRestraint.
793 
794 - Several functions have been added to helper in creating restraints. They are create_simple_connectivity_on_rigid_bodies(), create_simple_connectivity_on_molecules(), create_simple_distance(), create_simple_diameter(), create_simple_em_fit(), and create_simple_excluded_volume_on_rigid_bodies(). Functions load_em_density_map() and set_rigid_bodies() are added as well. See IMP::helper for more information.
795 
796 
797 - IMP::ScoreStates are further hidden. Namely, IMP::core::RigidBody constructs the needed score states internally, doing away with the need for the create functions (which have been done away with). In addition there are new decorators IMP::core::Cover and IMP::core::Centroid which manage their respective score states, further doing away with more create methods. The hierarchy-based create method for rigid bodies has been moved to the atom module and is called IMP::atom::rigid_body_setup_hierarchy().
798 
799 - IMP::em::read_map() will figure out the file type to use automatically from the extension. Other such methods should follow suite.
800 
801 - A class IMP::core::StructureSet has been added to provide a convenient way to manage solutions of optimization runs.
802 
803 - IMP::core::RigidBodyMover now takes its angle argument in radians and is documented as such.
804 
805 - IMP::core::ClosePairsScoreState and IMP::core::CloseBipartitePairsScoreState have been replaced by IMP::core::ClosePairContainer and IMP::core::CloseBipartitePairContainer. The score states have been marked as deprecated and are mostly still there, but are more limited than they were.
806 
807 - IMP::core::AllPairsPairContainer has been renamed to IMP::core::AllPairContainer and a bipartite version has been added.
808 
809 - All containers now keep track of what they looked like on the previous time step.
810 
811 - The IMP::core::ListPairContainer and IMP::core::ListSingletonContainer have had their interface reduced a bit in order to make implementing change tracking easier.
812 
813 - The set of exceptions has been simplified to go along with simplification of the checks. We now have IMP::InternalException and IMP::UsageException which are thrown by the respective checks (and the former is thrown by IMP_FAILURE()). When you want to to pass an exception back to Python use one of IMP::ValueException or IMP::IndexException or IMP::IOException. There is also a new macro, IMP_THROW() to be used when you want to throw an exception with a nice error message. Please look at code you have worked on and make sure that the exceptions make sense given the new clarifications.
814 
815 - An IMP::ModelException is now thrown when IMP::Particle attributes attain invalid values. This exception can be caught to recover from failures of IMP::core::ConjugateGradients or difficult restraints.
816 
817 ## Week of October 15
818 
819 - The set of functions needed to get meta information about IMP::Restraint and IMP::ScoreState objects has hopefully stabilized. Sorry about the thrashing. If you have restraints or score states not in SVN, see the corresponding IMP_RESTRAINT type macros, the IMP::Interaction class and base classes for what exactly you need to implement.
820 
821 ## Week of October 8
822 
823 - IMP::Interaction was modified to split IMP::interaction::get_used_particles() into IMP::interaction::get_read_particles() and IMP::interaction::get_write_particles(). This makes it possible to check for IMP::ScoreState objects being in an inconsistent order (so that the input to one is updated after it is read). Later, we will be able to automatically order the score states ensure that there are no problems.
824 
825 - IMP::atom::Hierarchy has been modifier to get rid of the IMP::atom::Hierarchy::Type. Instead, all types are contained in the decorators. IMP::atom::get_by_type() was also updated to reflect this.
826 
827 - IMP_check() and IMP_assert() have been
828 replaced with IMP_USAGE_CHECK() and IMP_INTERNAL_CHECK()
829 respectively. Hopefully this makes it clearer when to use
830 which. IMP::set_check_level() has been revised accordingly (to use
831 IMP::USAGE and IMP::USAGE_AND_INTERNAL). Also, IMP_failure() and
832 IMP_not_implemented() have been renamed to IMP_FAILURE() and
833 IMP_NOT_IMPLEMENTED() to make all macros be all caps. To update
834 external code,
835  - replace IMP.EXPENSIVE with IMP.USAGE_AND_INTERNAL
836  - replace IMP.CHEAP with IMP.USAGE
837  - replace IMP_assert and IMP_check as above.
838 
839 ## Week of October 1
840 - IMP::ScoreState and IMP::Restraint now better track the set of particles that they use and the dependencies between the particles. To this end, a new base class IMP::Interaction has been added with methods to get the inter-particle interactions and the set of particles used.
841 
842 - .cpp, .h, .py and SConscript files which contain the string "DO NOT COMMIT" will be rejected on checkin. Use this to make experimental or debugging changes you are making so you don't accidentally commit them.
843 
844 - The search module is no longer built by default and is not longer supported. It will get moved out of svn shortly.
845 
846 - IMP::ScoreState objects now also have IMP::Interacting::get_interacting_particles() methods. Soon IMP::PairScore and IMP::SingletonScore objects will too.
847 
848 ## Week of September 27
849 
850 - Some support for gather statistics has been added to IMP::Model (see IMP::Model::set_gather_statistics()). If requested, an API can be provided so that the individual data points can be returned to programs.
851 
852 - A class IMP::core::RigidBodyDistancePairScore has been added which provides accelerated computations for the distance between two rigid bodies.
853 
854 - The SWIG input files were renamed from swig.i to swig.i-in to aid in dependency tracking.
855 
856 - rigid bodies now have two sets of particles associated with them in the various places they are used. One is the IMP::core::RigidBodyMember particles, as was previously the case. The other is defined through a refiner (and could be the RigidBodyMember particles if a IMP::core::RigidMembersRefiner is used). The idea is that you sometimes want some particles which are transformed along with the rigid body, but which are not really part of the representation of the rigid body. Examples would include marker points or multiresolution representations of proteins.
857 
858 ## Week of September 20
859 
860 - all \c .i files have been renamed to start with \c IMP_ to help make dependency tracking easier.
861 
862 - static builds (builds with \c static=True) build both the static and dynamic libraries and just static executables
863 
864 - the \c profile build target has been removed as \c static=True does the important bit of what is needed. Just do \c linkflags=['-pg'] to add the needed link flag for \c gprof.
865 
866 - Restraints should now implement IMP::Restraint::unprotected_evaluate() instead of IMP::Restraint::evaluate(). The latter can now be called outside of IMP::Model::evaluate() and will return the right answer always.
867 
868 - \c bin/imppy.sh no longer works, as was previously warned. Use \c tools/imppy.sh instead. The bin directory will go away eventually, for now the script just tells you to use the tools version.
869 
870 ## Week of August 30
871 
872 - A new score IMP::core::BoundingBox3DSingletonScore has been added to keep particles
873  in a box.
874 
875 - str should now work better in Python when applied to IMP::Object objects and a few others. Objects
876  should use the new IMP_SHOWABLE and IMP_SHOWABLE_INLINE macros instead of declaring show
877  functions directory to make it so that they work in Python. It is now a bug to \c print or \c str
878  to not return something useful in Python for any IMP object.
879 
880 ## Week of August 23
881 
882 - each module now has a version, eg IMP::get_module_version_info() which is also returned by the various restraints of the module. The Python code checks that loaded libraries have the expected version numbers to try to provide better messages when runtime linking problems occur.
883 
884 - the various bounding_box methods have been renamed to ::get_bounding_box() to be more consistent with the other names.
885 
886 - Python documentation is now parsed by doxygen
887 
888 ## Week of August 15
889 
890 - The Object::get_version_info() result now includes the SVN revision info when building from SVN. This means you can use it to more easily tell exactly what a built version of IMP was built from.
891 
892 ## Week of August 8
893 
894 - Treatment of hetatoms in hierarchies (and hence pdbs) is now documented. See IMP::atom::Hierarchy. This involved broadening IMP::atom::Residue to handle ligands and having IMP::atom::Hierarchy::LIGAND, IMP::atom::Hierarchy::AMINOACID and IMP::atom::Hierarchy::NUCLEICACID be handled by IMP::atom::Residue.
895 
896 - IMP::atom::read_pdb() currently drops HOH ATOM records on the floor. We should add a IMP::atom::Water decorator if anyone wants to handle them (since they aren't atoms and don't have elements).
897 
898 - IMP::atom::Mass() has been added for decorating particles with mass. IMP::atom::Atom mass has been moved to that.
899 
900 - The scheme for generating atom types is now fully specified. See the IMP::atom::AtomType documentation.
901 
902 - IMP::atom::Atom and IMP::atom::Residue have been somewhat restructured. Some of the helper functions have been renamed to be more descriptive or removed (if they duplicated other easily accessible functionality). get_residue_index() has been removed (since you can just call get_index() on the return result of the IMP::atom::get_residue()). IMP::atom::get_chain() now returns an IMP::atom::Chain. root() was renamed to IMP::atom::get_root() for consistency.
903 
904 - SWIG files are now generated from a file called swig.i in the pyext directory. This file only has to contain module-specific commands such as lines to wrap headers and instantiate module templates. No more boilerplate.
905 
906 - The contents of bin are moving to tools since the two directories both contained assorted development tools
907  (not all of which were executable). For now, bin/imppy.sh will still work, but that will go away soon.
908 
909 - IMP::Decorator::show() now just takes the stream (it used to take a prefix also, which was often ignored).
910  Anyone who implemented a decorator will have to delete the second argument of their show function.
911 
912 ## Week of August 1
913 
914 - Containers now have methods IMP::SingletonContainer::evaluate(), IMP::PairContainer::evaluate(),
915  IMP::SingletonContainer::apply() and IMP::PairContainer::apply(). Using these can be significantly
916  faster than looping over the particles externally. Similarly, there is IMP::SingletonScore::evaluate(),
917  IMP::PairScore::evaluate(), IMP::SingletonModifier::apply(), IMP::PairModifier::apply().
918 
919 - The decorator functions have been renamed to IMP::Decorator::setup_particle(), IMP::Decorator::decorate_particle()
920  and IMP::Decorator::particle_is_instance(). A shell script \c tools/update-decorators is provided which
921  updates the names in all code in subdirectories of the current directory. Please backup things first.
922 
923 - a function, IMP::Restraint::safe_evaluate() has been added to be used
924  when requesting the value of restraints outside of the IMP::Model::evaluate().
925 
926 ## Week of July 18
927 
928 - The atom::BondPairContainer has been renamed to atom::BondPairFilter
929  since it wasn't a real container. A real atom::BondPairContainer has
930  been added.
931 
932 - Out of source builds mostly works. To do an out of source build, make a link
933  to \c scons_tools and \c SConstruct from the source directory in the build
934  directory and then add the line \command{repository="../path_to_src_dir"} in
935  the \c config.py in the build directory. Then running \c scons in the build
936  directory will put generated files there. This allows you to make both a
937  \c release and \c fast build off of the same set of sources. Proper docs
938  will be added soon.
939 
940 - The SConscript setup for a module has been greatly simplified so it should
941  be easy enough for users to maintain their own module and updates to the
942  build system shouldn't involve changes to user SConscripts any more. This does
943  mean that any non-svn modules need to be updated. Look at IMP.core or another
944  module and copy what is there. Again, proper docs will be added soon.
945 
946 
947 ## Week of July 11
948 
949 - Per Javi's suggestion, the conventions have been split into
950  "Coding conventions" and "API conventions" and linked from the main page.
951 
952 - benchmarks got moved to their own module, IMP::benchmark and will
953  now be tracked over time
954 
955 ## Week of July 4
956 
957 - FilteredListSingletonContainer and FilteredPairSingletonContainer
958  went away as there was too much hard to factor code shared with
959  IMP::core::ListSingletonContainer and IMP::core::PairSingletonContainer. The
960  IMP::core::ClosePairsScoreState is now handled by a list of filters
961  on the class itself. That is, do
962  IMP::core::ClosePairsScoreState::add_close_pair_filter() to add a
963  filter.
964 
965 - Hooks have been added to allow actions to be taken on an assertion
966  or check failure. See IMP::HandleFailure and its descendants for
967  more details.
968 
969 - IMP::display::CGOWriter has been much improved and now allows
970  provides lots of display-time control and easy building of
971  animations.
972 
973 - Generic optimizer states have been added
974  (IMP::core::SingletonsOptimizerState and
975  IMP::core::PairsOptimizerState).
976 
977 ## Week of June 15
978 
979 - box distributions (IMP::algebra::random_vector_on_box(),
980  IMP::algebra::random_vector_in_box()) take bounding boxes
981 
982 - IMP::em::DensityMap are now IMP::Object objects. Some current
983  allocations on the stack may need to be changed (since ref counted
984  objects can't be stack allocated).
985 
986 - IMP::em::DensityMap::Read() is now deprecated (as is
987  IMP::em::DensityMap::Write) as they violated a slew of
988  conventions. Use IMP::em::read_map() and IMP::em::write_map()
989  instead.
990 
991 
992 ## Week of June 8
993 
994 - new methods IMP::atom::add_radii() and IMP::atom::add_bonds();
995 
996 - Hierarchys to IMP::atom::Hierarchies
997 
998 - IMP::Particles has been changed to a new type. It no longer plays
999  quite so well with Python lists (you can't create one from an
1000  arbitrary Python list).
1001 
1002 - The containers for multiple decorators have been revised to make them
1003  more useful. See IMP::Decorators. You can now implicitly convert,
1004  for %example, a IMP::core::XYZs to an IMP::Particles (as well as use
1005  IMP::Decorators::get_particles()). All methods that take a
1006  collection of particles with a certain type now instead take a
1007  IMP::Decorators.
1008 
1009 - A method that has expectations about the particles it takes should
1010  take decorators rather than particles. That is, anything which
1011  expects IMP::core::XYZ particles, should take an IMP::core::XYZs
1012  instead of IMP::Particles.
1013 
1014 - IMP_LIST() now takes an extra argument (the name of the plural
1015  container)
1016 
1017 - In C++ IMP::Decorator objects are implicitly convertible to
1018  IMP::Particle* objects. Unfortunately, making this work in Python is
1019  problematic.
1020 
1021 - More methods take and return decorators rather than
1022  IMP::Particles. This means it is clearer what their expectations are
1023  and what they produce.
1024 
1025 ## Week of May 15
1026 
1027 - The IMP::helper module was added and various functionality moved
1028  there. See its documentation page for a list of what is there.
1029 
1030 - IMP::algebra::SphericalCoords was renamed to
1031  IMP::algebra::SphericalVector3D to be consistent with other algebra
1032  classes.
1033 
1034 - IMP::algebra::Rotation2D is now IMP::InvalidDefault to be consistent
1035  with the other similar classes.
1036 
1037 - We are trying to be a bit more specific about what is expected from
1038  a IMP::atom::Hierarchy. Towards that end, there are two methods for
1039  extracting a representation from a hierarchy,
1040  IMP::atom::get_detailed_representation(), and
1041  IMP::atom::get_simplified_representation()
1042 
1043 - A method has been added to produce a simplified copy of an
1044  IMP::atom::Hierarchy. The method is still a bit crude and could
1045  benefit from other people inspecting the output and saying how it
1046  should be improved.
1047 
1048 - There is now an IMP::algebra::BoundingBoxD (and
1049  IMP::algebra::BoundingBox3D) to track bounding boxes. There are
1050  methods like IMP::algebra::bound_box(Sphere3D) which gets the
1051  bounding box of a sphere.
1052 
1053 - Each IMP::Object has its own log level. If
1054  IMP::Object::set_log_level() is called with a value other than
1055  IMP::DEFAULT, the value is used for code within and called by the
1056  object. To support this in your objects you need to add
1057  IMP_OBJECT_LOG to the start of each major method. It has been added
1058  to some of the core objects.
1059 
1060 ## Week of May 8
1061 
1062 - IMP::atom::MolecularHierarchy has been renamed to
1063  IMP::atom::Hierarchy. Use \command{sed -i.old
1064  "s\#MolecularHierarchy\#Hierarchy\#g" my_files}
1065 
1066 - the long-deprecated Name(Decorator) was removed
1067 
1068 - The decorator suffix has been removed from decorator names.
1069 
1070 - Unit conventions have been established and documented for distance,
1071  energy, force and angle.
1072 
1073 - there is a function IMP::atom::create_protein() to create a coarse
1074  grained protein. The function is crude so far, but the idea is that
1075  we should have standard ways of generating such things so that
1076  everyone can benefit from best practices.
1077 
1078 - a class IMP::atom::BondGraph has been added so Boost.Graph
1079  algorithms can be used with bond graphs
1080 
1081 - the IMP-specific doxygen commands have been documented.
1082 
1083 - the doxygen commands unstable{name} and untested{name} have been
1084  added to doxygen to be used for marking methods and functions as
1085  likely to change or not well tested. Pass the name of the class or
1086  method as name.
1087 
1088 - bonds can now be added to molecular hierarchies automatically see
1089  IMP::atom::add_bonds()
1090 
1091 ## Week of May 1
1092 
1093 - There is a IMP::Rotation2D
1094 
1095 - You can now extract fixed Euler ZYZ angles from a
1096  IMP::Rotation3D. This code can serve as a template for extracting
1097  other Euler angle conventions if they are needed.
1098 
1099 - IMP::algebra::rotation_between_two_vectors() changed to
1100  IMP::algebra::rotation_taking_first_to_second() to make it clear
1101  which direction the rotation goes.
1102 
1103 ## Week of April 27
1104 
1105 - IMP::Decorator derived objects are now convertible to bool. So code
1106 like \code if (d) { blah } \endcode works.
1107 
1108 - IMP::Decorator::cast now returns a null decorator (\c
1109  DecoratorName()) rather than throwing an exception. This might make
1110  it actually useful.
1111 
1112 - New IMP::statistics module for statistical methods
1113  (e.g. clustering).
1114 
1115 ## Week of April 20
1116 
1117 - Support for taking advantage of rigid bodies in collision detection
1118  has been added via the IMP::core::RigidClosePairsFinder.
1119 
1120 - All IMP::core::ClosePairsScoreState-related functionality has been
1121  optimized and should be noticeably faster.
1122 
1123 - The EMBED project has been merged into the IMP::em module. Thus, it
1124  is no longer necessary to obtain EMBED separately to build IMP::em.
1125 
1126 - All C++ methods which work with std::istream or std::ostream
1127  arguments should now be usable from Python: any Python file-like
1128  object can be used.
1129 
1130 - The SWIG interface now ties in to the reference counting mechanism
1131  used in the C++ code, for all classes derived from
1132  IMP::RefCounted. So, for %example, deleting a Python object should
1133  not free the underlying C++ object if another C++ object holds a
1134  reference to it, and unrefing a C++ object should not delete it if a
1135  Python object still holds a reference to it. There should be no
1136  further need to manually keep references to Python objects to
1137  prevent things from being prematurely cleaned up. This should be
1138  transparent to the user; please report any problems.
1139 
1140 - IMP::atom::AtomDecorator no longer inherits from
1141  IMP::core::XYZDecorator.
1142 
1143 ## Week of April 13
1144 
1145 - The decorators IMP::atom::AtomDecorator, IMP::atom::ChainDecorator,
1146  IMP::atom::ResidueDecorator etc. now inherit from
1147  IMP::atom::MolecularHierarchyDecorator. This means they can set
1148  their IMP::atom::MolecularHierarchyDecorator::Type properly and
1149  saves a lot of typing in the common case. We can always make
1150  hierarchy decorators zero memory overhead if we find that the cost
1151  matters.
1152 
1153 - The base classes IMP::Comparable, IMP::NullDefault,
1154  IMP::ValidDefault and IMP::UninitializedDefault were added to
1155  document what the default constructors do for various objects in
1156  IMP.
1157 
1158 ## Week of April 6
1159 
1160 - IMP::Object now defines pure virtual methods
1161  IMP::Object::get_version_info() and IMP::Object::show(). All
1162  non-abstract classes which inherit from IMP::Object must implement
1163  them. Doing this means they are more uniformly implemented and are
1164  documented in one place.
1165 
1166 - IMP::Object is now reference counted (the only objects this changes
1167  are IMP::Model and IMP::Optimizer). As a result optimizers and
1168  IMP::Model objects must be created on the heap.
1169 
1170 - IMP::RefCounted has had its name changed and no longer inherits from
1171  IMP::Object. This makes it easier to use independently.
1172 
1173 - the docs for reference counting (IMP::RefCounted) and and IMP::Decorator
1174  have been improved
1175 
1176 - running doxygen produces no warnings. Let's try to keep it that
1177  way. Run \command{scons doc > /dev/null} before you check in large
1178  blocks of new code to make sure that you didn't introduce warnings.
1179 
1180 - Helper classes have been added to make SingletonModifier and
1181  PairModifier objects look like functors for C++ and Python. This
1182  makes it so you can use \c map in Python and \c for_each in C++. See
1183  IMP::SingletonFunctor and IMP::PairFunctor for more information.
1184 
1185 - The recent changes page has moved to a doxygen page to make it
1186  easier to link changes back to the relevant source code.
1187 
1188 ## Week of March 20:
1189 
1190 - create_rigid_body(), create_rigid_bodies(), create_centroid()
1191  etc. now require you to add the score state to the model
1192  yourself. Whether the rigid body is enforced or not can be
1193  controlled by adding or removing the returned IMP::ScoreState.
1194 
1195 - All previously deprecated code has been removed. This includes
1196  indexes in particles, the no-argument constructor for particles and
1197  the functionality that was moved from core to atom.
1198 
1199 - The IMP::core::RigidBodyDecorator and
1200  IMP::core::RigidMemberDecorator have been simplified by removing the
1201  traits classes. There is now only one type of rigid body and it
1202  stores the list of member particles internally.
1203 
1204 - IMP::core::NameDecorator has been deprecated as IMP::Particles have
1205  built in names (use IMP::Particle::get_name(),
1206  IMP::Particle::set_name()).
1207 
1208 - IMP::atom::ChainDecorator has been added
1209 
1210 ## Week of March 13:
1211 
1212 - A new IMP::display::Writer, IMP::display::ChimeraWriter has been
1213  added which exports geometry to native Chimera constructs. It
1214  probably should be used instead of bild and cmm for most purposes as
1215  it combines the best of both types.
1216 
1217 - compose() for transformation is now a free function so
1218  compose(rotation, vector) will work via promotions.
1219 
1220  - the geometry classes have been cleaned up and are now more
1221  consistent with IMP standards (and more likely to do what their
1222  names/comments suggest). See IMP::algebra::Sphere3D,
1223  IMP::algebra::Cylinder3D, IMP::algebra::Segment3D,
1224  IMP::algebra::Plane3D etc.
1225 
1226 - ParticleRefiner has been renamed to IMP::Refiner since the Particle
1227  was long and unnecessary.
def temporary_directory