Changelog

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

Unreleased

Added

• join_blocks, to join multiple blocks along samples or properties, with similar semantics to join

Changed

• join was rewritten and made faster. As part of this, we had to change the API: sort_samples is not an option anymore, and remove_tensor_name: bool was replaced by add_dimension: Optional[str]=None. You can set add_dimension="tensor" to get the old behavior of remove_tensor_name=False. The new default behavior is to not add extra dimensions, and fail if the resulting samples or properties are not unique.

Version 0.3.4 - 2025-07-28

Fixed

• insert_dimension no longer fails when called on empty TensorMap (#924)

Added

• equal_metadata now takes a check_gradients=True/False flag to enable or disable metadata checks on the gradients (#948)

Version 0.3.3 - 2025-04-25

Fixed

• Changed the implementation of std_over_samples and var_over_samples to reduce numerical noise issues when working with float32 (#908)

• Make sure the filtered TensorMap has keys in the same order as the filter in filter_blocks (#885)

• Make the code compatible with metatensor-torch v0.7.6

Version 0.3.2 - 2025-02-18

Fixed

• mean_over_samples was not computing the correct mean

Version 0.3.1 - 2025-02-04

Added

• Added a new filter_blocks function to remove some blocks from a TensorMap according to a selection (#823)

Fixed

• Fixed the sort function when dealing with empty blocks (#840)

Version 0.3.0 - 2024-10-29

Changed

• axis="components" can now be passed to metatensor.operations.rename_dimension (#763)

• split, split_block, slice and slice_block have been updated to accept a list of integer directly specifying which samples/properties to keep in the output. The corresponding parameter has been renamed to selection or selections (for slice and split respectively) (#772)

Version 0.2.4 - 2024-10-11

Changed

• block_from_array now takes optional parameters to specify the names of sample, component and property dimensions

Version 0.2.3 - 2024-08-28

Changed

• We now require Python >= 3.9

• slice and drop_blocks are now faster thanks to Labels.select

Version 0.2.2 - 2024-06-19

Fixed

• Fixed a bug in metatensor.torch.sort where the labels where not properly sorted (#647)

• Fix metatensor.abs when used with complex values (#553)

Version 0.2.1 - 2024-03-01

Changed

• Use torch.jit.script by default on all operations when using the TorchScript backend (i.e. metatensor.torch) (#504)

Version 0.2.0 - 2024-02-07

Changed

• join() operation now includes the sort_samples argument to control the sorting of samples in the output joined TensorMap. This defaults to False, opposite to the previous default behaviour.

Added

• detach() operation to detach all values in a TensorMap/TensorBlock from any computational graph

• requires_grad() operation to set the requires_grad of all values in a TensorMap/TensorBlock when storing data in torch Tensors.

Removed

• the to operation was removed. Similar functionality is now offered by TensorMap.to, TensorBlock.to, and the operations detach() and requires_grad().

Version 0.1.0 - 2023-10-11

Added

• Creation operations: empty_like(), ones_like(), zeros_like(), random_like(), block_from_array();

• Linear algebra: dot(), lstsq(), solve();

• Logic function: allclose(), equal(), equal_metadata();

• Manipulation operations: drop_blocks(), join(), manipulate dimension, one_hot(), remove_gradients(), samples reduction, slice(), split(), to();

• Mathematical functions: abs(), add(), divide(), multiply(), pow(), subtract();

• Set operations: unique_metadata(), sort();