Releases: ur-whitelab/hoomd-tf
Releases · ur-whitelab/hoomd-tf
v2.4.0
v2.3.0
New Features
- Can generate HOOMD snapshot bonds group based on mapped exclusion lists
Enhancements
- Made it possible to have multiple molecule features
Bug Fixes
- Fixed bug in compute_nlist with types not masking out of cut nlists
- Fixed PBC error on molecular features
- Fixed bug on calls to model with 2 inputs
v2.2.0
Breaking Changes
- Removed
sample_weight. Was not used by anyone and is an edge case.
New Features
- Can read molecules from GROAMCS topology
- Can compute one-hot encoding for different bead type interactions in CG systems
- Can generate mapped exclusion lists in CG systems
Enhancements
- Added jupyter notebook testing to continuous integration
iter_from_trajectorycan run in a user-defined time range- Revised to work for TF 2.4
- Can add exclusion lists when computing nlists using htf
v2.1.0
New Features
- Added standard layers to code
- Can compute molecular features for CG models
- Can compute CG matrix given a molecule and bead mappings
Enhancements
- Added error checks on mapping shapes
- Added optional training flag to change behavior during training
- Added ability to retrace compute if self changes
- Can output type or index when computing nlist
Bug Fixes
- Removed unused cuda events which could cause segfault on destruction
- Fixed simulation box in
iter_from_trajectory - Fixed
MDAnalysisuniverse based on atom groupselectionoption initer_from_trajectory
v1.0.2
v2.0.1
v2.0.0
New Features
- TensorFlow 2.0 now required
- Keras models replace computation graphs
- It is no longer required to save/load models for use
- Training is handled by Keras/TF functions now
- Getting values no longer requires save/load, instead everything is accessible in Python
Breaking Changes
- All code must be rewritten following new syntax
- feeddict is no longer supported
- graphbuilder is now SimModel, which must be subclassed
- Mol batching is now MolSimModel, which must be subclassed
- All graphbuilder methods (e.g., compute_rdf) are now functions
- Code that duplicates Keras functionality is removed:
- checkpoint management, model save/load
- saving values, computing means, other metrics
- saving values over time is now done by tensorboard or other TF/Keras approaches
- How forces are computed must be explicit and virials are no longer implicit (use modify_virial flag)
- You can no longer save tensors, instead output what you would like to save in your model code
- EDS is now layer (EDSLayer)
- Running from a trajectory is now done via a generator
Enhancements
- There is no more multiprocessing required
- Benchmarks are significantly faster
v1.0.1
v1.0
JOSS Review
Hoomd-TF has been published as a peer-reviewed article in the
Journal of Open Source Software (JOSS)
New Features
- Added pre-built molecular features
- Added MDAnalysis style selection for defining mapping operators
Enhancements
- Docs can now be built without HTF install
- mol batching performance is much better
- Simplified variable saving
- More example notebooks and reduced file sizes of example trajectories
- Supports dynamic boxes
- Better EDS bias naming
- Prevents accidentally computing forces from positions, instead of nlist
- Added guards against compiler mismatch
- Added sanity tests to prevent unsupported CPU/GPU modes
- Added benchmarking script
- Added check for nlist overflows on GPU
- Added check for mismatch on saving variables/tensors
- Fixed all compiler warnings
- Added Dockerfile for containerized use
Bug Fixes
- Empty tensorboard summaries no longer crash
- Prevented import issues with name clashes between packages and classes