v0.5.0

[0.5.0] - 2021-11-24

Changed

• Allow e3nn 0.4.*, which changes the default normalization of TensorProducts; this change should not affect typical NequIP networks
• Deployed are now frozen on load, rather than compile

Fixed

• load_deployed_model respects global JIT settings

[0.4.0] - not released

Added

• Support for e3nn's soft_one_hot_linspace as radial bases
• Support for parallel dataloader workers with dataloader_num_workers
• Optionally independently configure validation and training datasets
• Save dataset parameters along with processed data
• Gradient clipping
• Arbitrary atom type support
• Unified, modular model building and initialization architecture
• Added nequip-benchmark script for benchmarking and profiling models
• Add before option to SequentialGraphNetwork.insert
• Normalize total energy loss by the number of atoms via PerAtomLoss
• Model builder to initialize training from previous checkpoint
• Better error when instantiation fails
• Rename npz_keys to include_keys
• Allow user to register graph_fields, node_fields, and edge_fields via yaml
• Deployed models save the e3nn and torch versions they were created with

Changed

• Update example.yaml to use wandb by default, to only use 100 epochs of training, to set a very large batch logging frequency and to change Validation_loss to validation_loss
• Name processed datasets based on a hash of their parameters to ensure only valid cached data is used
• Do not use TensorFloat32 by default on Ampere GPUs until we understand it better
• No atomic numbers in networks
• dataset_energy_std/dataset_energy_mean to dataset_total_energy_*
• nequip.dynamics -> nequip.ase
• update example.yaml and full.yaml with better defaults, new loss function, and switched to toluene-ccsd(t) as example
  data
• use_sc defaults to True
• register_fields is now in nequip.data
• Default total energy scaling is changed from global mode to per species mode.
• Renamed trainable_global_rescale_scale to global_rescale_scale_trainble
• Renamed trainable_global_rescale_shift to global_rescale_shift_trainble
• Renamed PerSpeciesScaleShift_ to per_species_rescale
• Change default and allowed values of metrics_key from loss to validation_loss. The old default loss will no longer be accepted.
• Renamed per_species_rescale_trainable to per_species_rescale_scales_trainable and per_species_rescale_shifts_trainable

Fixed

• The first 20 epochs/calls of inference are no longer painfully slow for recompilation
• Set global options like TF32, dtype in nequip-evaluate
• Avoid possilbe race condition in caching of processed datasets across multiple training runs

Removed

• Removed allowed_species
• Removed --update-config; start a new training and load old state instead
• Removed dependency on pytorch_geometric
• nequip-train no longer prints the full config, which can be found in the training dir as config.yaml.
• nequip.datasets.AspirinDataset & nequip.datasets.WaterDataset
• Dependency on pytorch_scatter