# Changelog
All notable changes to this project will be documented in this file.

The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).

Most recent change on the top.

## Unreleased

## [0.19.1]

### Added
- docs page for the NequIP [foundation potentials](https://doi.org/10.48550/arXiv.2607.28461), with links to fine-tuning, compilation and the benchmarks reported in the paper, and citation instructions (`CITATION.bib`, `README`, docs) for using them
- `TorchProfilerCallback` for profiling training with `torch.profiler`
- `NEQUIP_NODE_PARALLEL_COMPILE` environment variable (`1`/`true`/`yes`/`y`) to serialize the first train-time compile within each node. Intended for large multi-node jobs, where every rank generating the same kernels at once against a shared cache directory on a parallel filesystem can crash or hang. Off by default.

### Fixed
- multi-rank checkpoint resume restoring tensors onto the GPU index recorded in the checkpoint rather than each rank's own device: `SimpleDDPStrategy` now moves the model to its device before restoring state (which also avoids `openequivariance` modules wedging when state is restored onto them beforehand), and loads model, optimizer, and EMA state directly onto that device
- excessive host memory use when resuming from a checkpoint: `nequip-train` held two full copies of the checkpoint in CPU memory at once (one to read hyperparameters and the run index, one from `load_from_checkpoint`); the first is now freed before the second load, and the second is loaded with `map_location="cpu"`
- https://github.com/mir-group/nequip/issues/612: `Expected all tensors to be on the same device` when loading a CUDA-built package on a CUDA-visible machine, seen as `nequip-package update` failing its own verification: e3nn's `CodeGenMixin` serializes generated `fx.GraphModule`s as opaque nested pickles whose storages carry an absolute device tag and are unreachable by `map_location`, so the Wigner-3j buffers stayed on `cuda:0` while the rest of the model loaded on CPU; the CPU-remapping context manager added for CPU-only hosts is now applied unconditionally
- silently NaN parameter gradients (and hence all-NaN metrics from the first optimization step) under train-time compile (`compile_mode: compile`) for frames with a zero-volume cell, e.g. non-periodic data, for which ASE supplies an all-zero cell: stress is `virial / volume`, and although it is unused by the loss, AOTAutograd bakes every output's backward into the joint graph and hands unused outputs a materialized zero tangent, so `0/0` produced a NaN that merged into the displacement gradient shared with forces and poisoned every parameter gradient (eager never schedules that branch, so it was unaffected); zero-volume frames now get NaN stress from a division that never sees a zero denominator
- `nequip-compile` failing on PyTorch 2.13 with `RuntimeError: PyTorch is checking whether allow_tf32 is enabled for cuDNN ...` for both `--device cuda` and `--device cpu` on machines with a CUDA-enabled build: 2.13 propagates `torch.backends.fp32_precision` to cuDNN, whose legacy `allow_tf32` flag cannot encode `"ieee"` and raises when `torch.export` snapshots it; the legacy flag is now kept in sync, leaving all `fp32_precision` values (and hence numerics) untouched
- train-time compile (`compile_mode: compile`) crashing on PyTorch 2.12 with `derivative for aten::silu_backward is not implemented`: 2.12 keeps `make_fx`'s decomposition table in a thread-local contextvar not seen by the autograd backward worker thread, so the `silu_backward` patch was skipped; tracing is now single-threaded on 2.12 to keep the table visible (no-op on other versions)
- silent wrong forces in the torch-sim integration for batched evaluation: `NequIPTorchSimCalc` used a cached system count, so a batch reusing the previous call's atomic numbers with a different number of systems (e.g. monatomic cells) used a stale layout; the system count now comes from the input state's cell batch dimension

## [0.19.0]

### Added
- `nequip-package list`: list the contents (with file sizes) of a packaged model file
- `nequip-package diff`: diff a file inside a package against its installed or a local version; useful to check for code drift between the packaged snapshot and the currently installed libraries
- `nequip-package update`: replace specific files inside a package and verify that model predictions are unchanged; intended for applying bug-fixes or compatibility patches without retraining
- `nequip-package modify`: apply one or more persistent model modifiers to a packaged model (e.g. adjust per-type energy scales/shifts, toggle force/stress output)
- per-species force loss weighting: `per_type_coeffs` key on `MetricsManager` metrics, and `per_type_forces_coeffs` kwarg on `EnergyForceLoss` and `EnergyForceStressLoss` for emphasizing some atom types over others in the forces loss (weighted-mean aggregation over per-type MSEs)
- AOTI-packaged models now record the custom-op libraries their GPU kernels require (e.g. `openequivariance`) and import them automatically at load time, so packaged models with custom GPU kernels can be loaded in plain Python without first manually importing the kernel library (only applies to models packaged with this version or later; previously packaged models do not carry this metadata and still require manual import)

### Changed

- [Breaking] `MetricsManager` internal refactor: the `metrics` attribute (a `dict` of `dict`s) has been replaced by `entries` (a `dict` of `MetricEntry` dataclasses). Existing checkpoints are incompatible. Callbacks and code that accessed `loss.metrics[name]["coeff"]` must be updated to `loss.entries[name].coeff`; affected built-in callbacks: `SoftAdaptCallback`, `LossCoeffScheduler`, `LossCoeffMonitor`.
- [Breaking] minimum `lightning` requirement raised to `>=2.6` (from `>=2.0`); environments pinned to older lightning must upgrade. Checkpoint loading now passes `weights_only=False` to accommodate lightning 2.6's new `weights_only=True` default

### Fixed

- https://github.com/mir-group/nequip/issues/601
- silent wrong forces/stress in the torch-sim integration: non-contiguous input tensors (e.g. the transposed `row_vector_cell` from `torch_sim.concatenate_states`/`BinningAutoBatcher`, or stride-0 expanded PBC) were read by AOTI models using assumed strides; all model inputs are now made contiguous
- double-backward overflow in large models: `silu_backward` is now evaluated via the `torch.sigmoid` primitive (instead of the `1/(1+exp(-x))` sigmoid expansion) through a patched AOTI decomposition table

## [0.18.0]

### Added
- XPU support for training and inference on Aurora
- Muon optimizer support
- validation-time XYZ file writer callback support
- `MappedFieldModifier` for mapped field transformations

### Changed
- train-time compilation now uses `fullgraph=False`

### Fixed
- https://github.com/mir-group/nequip/issues/592
- LAMMPS ML-IAP with type-dependent cutoffs bug

## [0.17.1]

### Added
- `dataset` as a built-in `AtomicDataDict` graph/long field (`AtomicDataDict.DATASET_KEY`) and `nequip.data.transforms.DatasetIndexTransform` for multi-dataset workflows

### Fixed
- compatibility with `alchemiops` v0.3.0 neighborlist

## [0.17.0]

### Added
- train-time compilation support for models with different input-output key signatures - enables holding multiple compiled models with different signatures
- documentation section on known PyTorch version issues
- convenience resolver `${cutoff_radius_from_package:${model_path}}` to extract cutoff radius for fine-tuning directly in config file
- `PresetNequIPGNNModel` model builder with `S`/`M`/`L`/`XL` architecture presets and explicit-argument override behavior
- docs linking to updated OpenMM integration
- `alchemiops` neighborlist backend support

### Changed
- overhauled `NonperiodicCellTransform` with improved handling
- models compiled for training now produce full set of eager output dict entries
- `NequIPCalculator` canonical import path is now `nequip.integrations.ase.NequIPCalculator`; `nequip.ase.NequIPCalculator` remains as a deprecated compatibility shim that emits a `FutureWarning`
- [Breaking] `NEQUIP_NL` environment variable no longer controls neighborlist backend selection. Neighborlist backend can be set explicitly via `NeighborListTransform(..., backend="...")`; default behavior is unchanged (`matscipy`).
- Added `register_neighborlist_backend` to support extensible neighborlist backend registration.

### Fixed
- PyTorch 2.10.0 tracing error for `silu_backward`
- PyTorch 2.10.0 AOTI codecache error
- https://github.com/mir-group/nequip/issues/572

## [0.16.3]

### Added
- Zenodo archiving

### Fixed
- syntax in `CITATION.cff`, verified with `cffconvert --validate -i CITATION.cff`

## [0.16.2]

### Added
- possibility of configuring nonlinear readout MLP for NequIP GNN models

### Changed
- bumped minimum e3nn version to 0.5.9

### Fixed
- added missing `Ru` element in the `TM23DataModule`
- https://github.com/mir-group/nequip/issues/578

## [0.16.1]

### Fixed
- Made torch-sim integration optional: removed `torchsim` import from `nequip.integrations.__init__.py` to prevent import errors when `torch_sim` is not installed. This allows other integration tools (e.g., `nequip-prepare-lmp-mliap`) to work without requiring `torch_sim` as a dependency.

## [0.16.0]

### Added
- users can specify having irreps of different multiplicites in `NequIPGNNModel` by providing `num_features` that is a list of `l_max + 1` features. E.g. for `l_max=2` and `parity=False`, `num_features=[5, 2, 7]` refers to `5x0e`, `2x1o` and `7x2e` features (see `configs/tutorial.yaml` for an example)
- users can specify `type_embed_num_features` as a separate hyperparameter to control the number of features in the type embedding layer (defaults to `num_features[0]`)
- batched AOTI inference
- per-edge-type cutoff can now lead to cost reduction in the LAMMPS ML-IAP interface
- optional `--constant-fold` acceleration argument for `nequip-compile --mode aotinductor` that can provide small speed ups for PyTorch >= 2.8 (may fail with some models, please open issues if that such instances are encountered)
- OpenEquivariance is now compatible with AOTInductor compilation for ASE inference with PyTorch >= 2.9
- better handling of datasets where stresses are partially populated, e.g. `AddNaNStressTransform` added to add NaN stress tensors for structures without stress data
- automatic caching for nequip.net models
- per-type average number of neighbors normalization option
- torch-sim integration with `NequIPTorchSimCalc` for batched GPU evaluation of compiled NequIP and Allegro models

### Removed
- [Breaking] `NequIPGNNEnergyModel` has been removed. Energy-only models can be constructed by using `NequIPGNNModel` with a new argument `do_derivatives=False`

### Changed
- Simplified metrics manager wrappers now include maximum absolute error metrics
- [Breaking] `ChemicalSpeciesToAtomTypeMapper` and `NequIPCalculator` API changed for safety and better UI with regards to atom types when using pretrained models
- [Breaking] raise minimum PyTorch version for OpenEquivariance to PyTorch >= 2.7 (previously >= 2.4)


## [0.15.0]

### Added
- Possible to now use `ChainedScheduler` and `SequentialLR` for LR scheduling
- Auto-download capabilities for pretrained potentials from [nequip.net](https://www.nequip.net/)

### Changed
- Loosen LAMMPS ML-IAP constraints of using absolute paths and ensuring that the original checkpoint or package file is present in the same exact location during ML-IAP runtime

### Fixed
- OpenEquivariance accelerated compiled training with PyTorch 2.8
- `NequIPCalculator` usage restrictions are loosened, e.g. when mixing ASE calculators


## [0.14.0]

### Added
- CuEquivariance acceleration for NequIP GNN models
- ScheduleFreeLightningModule: New LightningModule subclass that integrates Facebook's Schedule-Free optimizer to enable schedule-free training in NequIP.
- `TestTimeXYZFileWriter` now accepts `extra_fields` argument
- Add `WaterDataModule` and `COLLDataModule`

### Changed
- NequIP GNN models now exploit ghost node truncation in LAMMPS ML-IAP (optimization)
- Better ML-IAP error handling
- [Breaking] Disallow list input for per-type energy scales and shifts (only scalars or dicts allowed)
- [Breaking] Ignore `global_options` in config file; `TF32Scheduler` is the only way to configure TF32 settings during training

### Fixed
- `nequip-compile --mode aotinductor` failure with PyTorch 2.8 (because of constant-folding -- so AOTI constant-folding is always off now)


## [0.13.0]

### Added
- LAMMPS ML-IAP integration
- `LinearLossCoefficientScheduler` callback
- `TF32Scheduler` callback

### Changed
- `EMTTestDataset` now uses orthorhombic cells


## [0.12.1]

### Added
- `MaximumAbsoluteError` metric

### Changed
- Update `SoftAdapt` callback to weight loss coefficient updates by the chosen (initial) loss coefficients.

### Fixed
- `PerTypeScaleShift` model modifier shape bug
- `WandbWatch` callback typo in docstring
- Broken links to get `fcu.xyz` example dataset for tutorials/examples
- `nequip-package` will now always pick up OEQ file even if OEQ is not installed at package time


## [0.12.0]

### Added
- `SAM23DataModule`: Specialized datamodule for Samsung's SAMD23 dataset (HfO and SiN systems), with automatic download, extraction, and optional OOD test set support.

### Fixed
- Train-time compile compatibility of OpenEquivariance accelerated NequIP models

### Removed
- [Breaking] Python 3.9 support has been removed


## [0.11.1]

### Fixed
- Fixed per-edge-type cutoff metadata parsing when loading compiled models

### Changed
- Change to make train-time compile works with PyTorch 2.8.0 -- train-time compile won't work with PyTorch 2.8.0 and NequIP versions before v0.11.1


## [0.11.0]

### Added
- Energy-only training support: `NequIPGNNEnergyModel`, `EnergyOnlyDataStatisticsManager`, `EnergyOnlyLoss`, and `EnergyOnlyMetrics` for datasets without force labels
- Per-edge-type cutoff handling in ASE Calculator

### Removed
- [Future] Python 3.9 support will be removed in the coming releases. For now, users will be warned to upgrade if using NequIP with Python 3.9.

### Changed
- [Breaking] Hide `NequIPCalculator.from_checkpoint_model` and `NequIPCalculator.from_packaged_model` methods by making them private; users should only ever use `NequIPCalculator.from_compiled_model`


## [0.10.0]

### Changed
- Restructured user guide docs

### Added
- OpenEquivariance acceleration for NequIP GNN models
- `SortedNeighborListTransform` for sorted neighborlists with permutation indices to get a sorted transpose
- Per-edge-type cutoff-aware neighborlist transform


## [0.9.1]

### Changed
- Revamped docs


## [0.9.0]

### Changed
- [Breaking] `nequip-compile` CLI: `--input-path` and `--output-path` are now positional arguments instead of flags
- [Breaking] `nequip-package build` CLI: `--ckpt-path` and `--output-path` are now positional arguments instead of flags


## [0.8.0]

### Added
- MD22 datamodule
- `PerTypeScaleShift` model modifier to enable changing the per-type atomic energy scales and shifts of a pretrained model for fine-tuning

### Changed
- [Breaking] packaged model file metadata format: NOTE that packaged models before this version will no longer be compatible, and models must be repackaged from checkpoints
- [Breaking] `nequip-package` API: users must now specify `nequip-package build` to construct packaged model files or `nequip-package info` to inspect the metadata of packaged model files
- `per_type_energy_scales` and `per_type_energy_shifts` are expected to be in dict format; warnings will be thrown if they are provided as lists, and this will become errors in later major releases
- [Breaking] `InteractionBlock` of NequIP model refactored such that checkpoint files from previous versions will not work with this version

### Removed
- [Breaking] train-time TorchScript


## [0.7.1]

### Fixed
- Update outdated descriptions for `parity_plot.py`

### Added
- Increase docs coverage
- `StressSignFlipTransform` as a data transform to flip the sign of datasets that come with the opposite stress sign convention as used in the NequIP framework
- `int_div` and `int_mul` resolvers for integer arithmetic in config files, e.g. `half_width: ${int_div:${width},2}`

### Changed
- Renamed `examples` -> `misc` directory


## [0.7.0]
A major backwards-incompatible update with breaking changes throughout the code.


## [0.6.2] - 2025-3-22

### Fixed
 - Fixed early stopping bug
 - Fixed PyTorch version bug

## [0.6.1] - 2024-7-9
### 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`


## [0.6.0] - 2024-5-10
### 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`

## [0.5.6] - 2022-12-19
### 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)

## [0.5.5] - 2022-06-20
### 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)

## [0.5.4] - 2022-04-12
### 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)

## [0.5.3] - 2022-02-23
### Added
- `nequip-evaluate --repeat` option
- Report number of weights to wandb

### Changed
- defaults and commments in example.yaml and full.yaml, in particular longer default training and correct comment for E:F-weighting
- better metrics config in example.yaml and full.yaml, in particular will total F-MAE/F-RMSE instead of mean over per-species
- default value for `report_init_validation` is now `True`
- `all_*_*` metrics rename to -> `psavg_*_*`
- `avg_num_neighbors` default `None` -> `auto`

### Fixed
- error if both per-species and global shift are used together


## [0.5.2] - 2022-02-04
### Added
- Model builders may now process only the configuration
- Allow irreps to optionally be specified through the simplified keys `l_max`, `parity`, and `num_features`
- `wandb.watch` via `wandb_watch` option
- Allow polynomial cutoff _p_ values besides 6.0
- `nequip-evaluate` now sets a default `r_max` taken from the model for the dataset config
- Support multiple rescale layers in trainer
- `AtomicData.to_ase` supports arbitrary fields
- `nequip-evaluate` can now output arbitrary fields to an XYZ file
- `nequip-evaluate` reports which frame in the original dataset was used as input for each output frame

### Changed
- `minimal.yaml`, `minimal_eng.yaml`, and `example.yaml` now use the simplified irreps options `l_max`, `parity`, and `num_features`
- Default value for `resnet` is now `False`

### Fixed
- Handle one of `per_species_shifts`/`scales` being `null` when the other is a dataset statistc
- `include_frames` now works with ASE datasets
- no training data labels in input_data
- Average number of neighbors no longer crashes sometimes when not all nodes have neighbors (small cutoffs)
- Handle field registrations correctly in `nequip-evaluate`

### Removed
- `compile_model`

## [0.5.1] - 2022-01-13
### Added
- `NequIPCalculator` can now be built via a `nequip_calculator()` function. This adds a minimal compatibility with [vibes](https://gitlab.com/vibes-developers/vibes/)
- Added `avg_num_neighbors: auto` option
- Asynchronous IO: during training, models are written asynchronously. Enable this with environment variable `NEQUIP_ASYNC_IO=true`.
- `dataset_seed` to separately control randomness used to select training data (and their order).
- The types may now be specified with a simpler `chemical_symbols` option
- Equivariance testing reports per-field errors
- `--equivariance-test n` tests equivariance on `n` frames from the training dataset

### Changed
- All fields now have consistant [N, dim] shaping
- Changed default `seed` and `dataset_seed` in example YAMLs
- Equivariance testing can only use training frames now

### Fixed
- Equivariance testing no longer unintentionally skips translation
- Correct cat dim for all registered per-graph fields
- `PerSpeciesScaleShift` now correctly outputs when scales, but not shifts, are enabled— previously it was broken and would only output updated values when both were enabled.
- `nequip-evaluate` outputs correct species to the `extxyz` file when a chemical symbol <-> type mapping exists for the test dataset

## [0.5.0] - 2021-11-24
### Changed
- Allow e3nn 0.4.*, which changes the default normalization of `TensorProduct`s; 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`

## [0.3.3] - 2021-08-11
### Added
- `to_ase` method in `AtomicData.py` to convert `AtomicData` object to (list of) `ase.Atoms` object(s)
- `SequentialGraphNetwork` now has insertion methods
- `nn.SaveForOutput`
- `nequip-evaluate` command for evaluating (metrics on) trained models
- `AtomicData.from_ase` now catches `energy`/`energies` arrays

### Changed
- Nonlinearities now specified with `e` and `o` instead of `1` and `-1`
- Update interfaces for `torch_geometric` 1.7.1 and `e3nn` 0.3.3
- `nonlinearity_scalars` now also affects the nonlinearity used in the radial net of `InteractionBlock`
- Cleaned up naming of initializers

### Fixed
- Fix specifying nonlinearities when wandb enabled
- `Final` backport for <3.8 compatability
- Fixed `nequip-*` commands when using `pip install`
- Default models rescale per-atom energies, and not just total
- Fixed Python <3.8 backward compatability with `atomic_save`

## [0.3.2] - 2021-06-09
### Added
- Option for which nonlinearities to use
- Option to save models every *n* epochs in training
- Option to specify optimization defaults for `e3nn`

### Fixed
- Using `wandb` no longer breaks the inclusion of special objects like callables in configs

## [0.3.1]
### Fixed
- `iepoch` is no longer off-by-one when restarting a training run that hit `max_epochs`
- Builders, and not just sub-builders, use the class name as a default prefix
### Added
- `early_stopping_xxx` arguments added to enable early stop for platued values or values that out of lower/upper bounds.

## [0.3.0] - 2021-05-07
### Added
- Sub-builders can be skipped in `instantiate` by setting them to `None`
- More flexible model initialization
- Add MD w/ Nequip-ASE-calculator + run-MD script w/ custom Nose-Hoover

### Changed
- PBC must be explicit if a cell is provided
- Training now uses atomic file writes to avoid corruption if interupted
- `feature_embedding` renamed to `chemical_embedding` in default models

### Fixed
- `BesselBasis` now works on GPU when `trainable=False`
- Dataset `extra_fixed_fields` are now added even if `get_data()` returns `AtomicData` objects

## [0.2.1] - 2021-05-03
### Fixed
- `load_deployed_model` now correctly loads all metadata

## [0.2.0] - 2021-04-30