Changelog¶

All notable changes to metatensor-learn are documented here, following the keep a changelog format. This project follows Semantic Versioning.

Unreleased ¶

Added metatensor.learn.nn.EquivariantTransformation to apply any torch.nn.Module to invariants computed from the norm over components of covariant blocks. The transformed invariants are then elementwise multiplied back to the covariant blocks. For invariant blocks, the torch.nn.Module is applied as is (#744)

metatensor.learn.nn modules InvariantTanh, InvariantSiLU, InvariantReLU, InvariantLayerNorm, and EquivariantLinear have removed and replaced parameter. invariant_key_idxs is replaced by invariant_keys, a Labels object that selects for invariant blocks.
metatensor.learn.nn modules LayerNorm, InvariantLayerNorm, Linear, and EquivariantLinear have altered accepted types for certain parameters. Parameters eps, elementwise_affine, bias, and mean for the layer norm modules, and bias for the linear modules are affected. Previously these could be passed as list, but now can only be passed as a single value. For greater control over modules applied to individual blocks, users are encouraged to use the ModuleMap module from metatensor.learn.nn.

We now require Python >= 3.9
Dataset and DataLoader can now handle fields with a name which is not a valid Python identifier.

Added torch-style activation function module maps to metatensor.learn.nn: ReLU, InvariantReLU, SiLU, and InvariantSiLU (#597)
Added torch-style neural network module maps to metatensor.learn.nn: LayerNorm, InvariantLayerNorm, EquivariantLinear, Sequential, Tanh, and InvariantTanh (#513)

metatensor.learn.nn modules LayerNorm and InvariantLayerNorm now applies sample-independent transformations to input tensors.
Set correct device for output of when torch default device is different than input device (#595)

metatensor-learn is no longer re-exported from metatensor and metatensor.torch, all functions are still available inside metatensor.learn and metatensor.torch.learn.

Pluralization removed for special kwarg sample_ids of IndexedDataset -> sample_id, and provided collate functions group and group_and_join updated accordingly.

ModuleMap and Linear modules, following torch.nn.ModuleDict and torch.nn.Linear in PyTorch but adapted for TensorMap’s (#427)
Dataset and DataLoader facilities, following the corresponding classes in PyTorch (#428)