Nequip

Fixed
- Fixed early stopping bug
- Fixed PyTorch version bug

Added
- add support for equivariance testing of arbitrary Cartesian tensor outputs
- [Breaking] use entry points for `nequip.extension`s (e.g. for field registration)
- alternate neighborlist support enabled with `NEQUIP_NL` environment variable, which can be set to `ase` (default), `matscipy` or `vesin`
- Allow `n_train` and `n_val` to be specified as percentages of datasets.
- Only attempt training restart if `trainer.pth` file present (prevents unnecessary crashes due to file-not-found errors in some cases)

Changed
- [Breaking] `NEQUIP_MATSCIPY_NL` environment variable no longer supported

Fixed
- Fixed `flake8` install location in `pre-commit-config.yaml`

Added
- add Tensorboard as logger option
- [Breaking] Refactor overall model logic into `GraphModel` top-level module
- [Breaking] Added `model_dtype`
- `BATCH_PTR_KEY` in `AtomicDataDict`
- `AtomicInMemoryDataset.rdf()` and `examples/rdf.py`
- `type_to_chemical_symbol`
- Pair potential terms
- `nequip-evaluate --output-fields-from-original-dataset`
- Error (or warn) on unused options in YAML that likely indicate typos
- `dataset_*_absmax` statistics option
- `HDF5Dataset` (227)
- `include_file_as_baseline_config` for simple modifications of existing configs
- `nequip-deploy --using-dataset` to support data-dependent deployment steps
- Support for Gaussian Mixture Model uncertainty quantification (https://doi.org/10.1063/5.0136574)
- `start_of_epoch_callbacks`
- `nequip.train.callbacks.loss_schedule.SimpleLossSchedule` for changing the loss coefficients at specified epochs
- `nequip-deploy build --checkpoint` and `--override` to avoid many largely duplicated YAML files
- matscipy neighborlist support enabled with `NEQUIP_MATSCIPY_NL` environment variable

Changed
- Always require explicit `seed`
- [Breaking] Set `dataset_seed` to `seed` if it is not explicitly provided
- Don't log as often by default
- [Breaking] Default nonlinearities are `silu` (`e`) and `tanh` (`o`)
- Will not reproduce previous versions' data shuffling order (for all practical purposes this does not matter, the `shuffle` option is unchanged)
- [Breaking] `default_dtype` defaults to `float64` (`model_dtype` default `float32`, `allow_tf32: true` by default--- see https://arxiv.org/abs/2304.10061)
- `nequip-benchmark` now only uses `--n-data` frames to build the model
- [Breaking] By default models now use `StressForceOutput`, not `ForceOutput`
- Added `edge_energy` to `ALL_ENERGY_KEYS` subjecting it to global rescale

Fixed
- Work with `wandb>=0.13.8`
- Better error for standard deviation with too few data
- `load_model_state` GPU -> CPU
- No negative volumes in rare cases

Removed
- [Breaking] `fixed_fields` machinery (`npz_fixed_field_keys` is still supported, but through a more straightforward implementation)
- Default run name/WandB project name of `NequIP`, they must now always be provided explicitly
- [Breaking] Removed `_params` as an allowable subconfiguration suffix (i.e. instead of `optimizer_params` now only `optimizer_kwargs` is valid, not both)
- [Breaking] Removed `per_species_rescale_arguments_in_dataset_units`

Added
- sklearn dependency removed
- `nequip-benchmark` and `nequip-train` report number of weights and number of trainable weights
- `nequip-benchmark --no-compile` and `--verbose` and `--memory-summary`
- `nequip-benchmark --pdb` for debugging model (builder) errors
- More information in `nequip-deploy info`
- GPU OOM offloading mode

Changed
- Minimum e3nn is now 0.4.4
- `--equivariance-test` now prints much more information, especially when there is a failure

Fixed
- Git utilities when installed as ZIPed `.egg` (264)

Added
- BETA! Support for stress in training and inference
- `EMTTestDataset` for quick synthetic fake PBC data
- multiprocessing for ASE dataset loading/processing
- `nequip-benchmark` times dataset loading, model creation, and compilation
- `validation_batch_size`
- support multiple metrics on same field with different `functional`s
- allow custom metrics names
- allow `e3nn==0.5.0`
- `--verbose` option to `nequip-deploy`
- print data statistics in `nequip-benchmark`
- `normalized_sum` reduction in `AtomwiseReduce`

Changed
- abbreviate `node_features`->`h` in loss titles
- failure of permutation equivariance tests no longer short-circuts o3 equivariance tests
- `NequIPCalculator` now stores all relevant properties computed by the model regardless of requested `properties`, and does not try to access those not computed by the model, allowing models that only compute energy or forces but not both

Fixed
- Equivariance testing correctly handles output cells
- Equivariance testing correctly handles one-node or one-edge data
- `report_init_validation` now runs on validation set instead of training set
- crash when unable to find `os.sched_getaffinity` on some systems
- don't incorrectly log per-species scales/shifts when loading model (such as for deployment)
- `nequip-benchmark` now picks data frames deterministically
- useful error message for `metrics_key: training_*` with `report_init_validation: True` (213)

Added
- `NequIPCalculator` now handles per-atom energies
- Added `initial_model_state_strict` YAML option
- `load_model_state` builder
- fusion strategy support
- `cumulative_wall` for early stopping
- Deploy model from YAML file directly

Changed
- Disallow PyTorch 1.9, which has some JIT bugs.
- `nequip-deploy build` now requires `--train-dir` option when specifying the training session
- Minimum Python version is now 3.7

Fixed
- Better error in `Dataset.statistics` when field is missing
- `NequIPCalculator` now outputs energy as scalar rather than `(1, 1)` array
- `dataset: ase` now treats automatically adds `key_mapping` keys to `include_keys`, which is consistant with the npz dataset
- fixed reloading models with `per_species_rescale_scales/shifts` set to `null`/`None`
- graceful exit for `-n 0` in `nequip-benchmark`
- Strictly correct CSV headers for metrics (198)