v0.0.3

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.