Skip to content

Latest commit

 

History

History
160 lines (120 loc) · 6.24 KB

CHANGELOG.md

File metadata and controls

160 lines (120 loc) · 6.24 KB
Raw

Changelog

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.1.1] - 2023-07-13

Fixed

  • Bayesian optimization for sigma hyperparameter optimization changed how to specify gp_params.
  • Pydantic had API changes that affected qcelemental, so we have to specify pydantic<2.0.0 for now.

[0.1.0] - 2023-04-20

Added

  • Iterative solver for GDML training on larger systems (changes up to v1.0.0).
  • Additional documentation on obtaining many-body data.
  • Some many-body expansion logging (for debugging).
  • Logger documentation.
  • Function to change mbGDML log levels.
  • Ability of ASE calculator to update periodic cell on mbePredict object.
  • Preliminary virial stress algorithms and stress module.
  • Switching function module for alchemical scaling.
  • Documentation example of an ASE optimization under periodic boundary conditions.
  • Structure generation module using packmol.
  • Provide explicit example of model comp ids in documentation.

Changed

  • GDML solvers are put into their own modules.
  • Updated prediction set documentation.
  • Unified use of ray by always having use_ray option and defaulting n_workers to 1.

Fixed

  • utils.center_structures sometimes repeated the array incorrectly.
  • Explicitly use 3DMol v1.8.0 for documentation.

[0.0.4] - 2022-12-12

Added

  • mbgdml.descriptors.Criteria class for setting up structure descriptors and cutoffs for models.
  • Model analysis with Matérn covariance function.
  • mbe_contrib tests.
  • Radial distribution function analysis
  • Periodic many-body expansions with the minimum image convention.
  • Many-body alchemical parameter.

Changed

  • Unified use of Z, z, R, and r.
  • dataSet to DataSet
  • Switch to qcelemental for atom properties.
  • Do not restrict sigma_bounds for Bayesian optimization after initial grid search.
  • mbgdml.criteria is now mbgdml.descriptors with a modular handling of descriptors and cutoffs.
  • Provide changeable Z, R, E, and F keys for loading npz data sets.
  • Modularize the error and loss calculations.
  • Reduce regularization strength to 1e-10 instead of 1e-15.
  • Split predict module into models and predictors.
  • Install bayesian-optimization from git until scipy bug is fixed.
  • Merged sGDML changes up to v0.5.4 (124de3dd8d46a0622bd10c3b4ab033a00dbd3c27).
  • Predict times are logged at debug level.
  • Ray must be initialized outside of mbePredict class.
  • Use n_workers instead of n_cores in mbePredict.

Fixed

  • mbe_worker with heterogeneous n-body structures (e.g., solute+solvent).
  • Custom todict method for ASE calculator. Fixes attached ASE trajectory in reading entity_ids.
  • Store ASE Atoms object to avoid recalculating energies and forces in ASE calculator.
  • Doc references to respective SchNet functions.

Removed

  • Dependency on natsort.
  • mbgdml.data.mbModel was adsorbed into mbgdml.models.gdmlModel.
  • structureSets and sampling for data sets are no longer supported and subsequently removed. This functionality was incorporated into reptar.

[0.0.3] - 2022-07-25

Added

  • SchNetPack prediction capabilities.
  • GAP prediction capabilities.
  • Training loss function that includes a weighted energy RMSE component.
  • Require integration constant evaluation option regardless of performance.
  • Initial grid for Bayesian optimization to guide sigma_bounds.
  • Ability to keep all trained models instead of just the best one.
  • Log parallel optimization.
  • Plot Gaussian process from hyperparameter Bayesian optimization.
  • Plot cluster losses and population histogram using matplotlib.
  • Option to use a sequential reduction optimizer for Bayesian optimization.
  • Specify Gaussian process keyword arguments for the final iterative training task.

Changed

  • Removing md module in favor of having an interfaces module.
  • Storage of n-body energies and forces in predict sets.
  • Redesigned predict methods and parallelized with ray.
  • Included a many-body expansion, mbe, module to handle n-body energy and force predictions.
  • Updated API documentation tree.
  • Elements logging in tasks and models are condensed (i.e., no spaces).
  • Default gp_params for Bayesian optimization.
  • MD5 hashes are no longer stored in bytes.
  • Do not include training set in any problematic clustering. Training structures are not included in dataset clustering or plots.
  • Training JSON to training.json instead of log.json.
  • Iterative training task directory names to state the training set size.

Fixed

  • Added missed torchtools for GDML.
  • model0 was not working with iterative training.
  • Iterative training would randomly sample every training set.

Removed

  • No longer can make many-body dataset with model predictions (with create_mb_from_models).
  • e_f_contributions was replaced by the mbe module.

[0.0.2] - 2022-06-08

Added

  • Iterative training procedure by finding problematic structures.
  • Bayesian optimization for hyperparameter search.
  • Basic logging capabilities.
  • Write JSON file after training with useful information.
  • Specify validation structures when training.

Changed

  • Sort md5_data keys for consistency.
  • Renamed add_pes_data to add_pes_json
  • asdict is now a method instead of a property.
  • Removed sGDML dependency.
  • Use relative imports.
  • Hyperparameter grid search in mbGDMLTrain class.
  • Moved sGDML modified training routines to _train.py.
  • Changed Rset_md5 to r_prov_ids and Rset_info to r_prov_specs.
  • Improved the write_xyz and string_coords functions.
  • comp_ids is now a 1D array where the index of the label is the entity_id.

Fixed

  • Grammar and typos in documentation.
  • Address Sphinx documentation warnings and errors.
  • Only deploy documentation on keithgroup repo.
  • Correct dataSet Rset_info documentation.

Removed

  • qc module. This does not belong in this package.

[0.0.1] - 2022-02-24

  • Initial release!