TensorLy Files

We're happy to release TensorLy 0.10.0! This release brings new PARAFAC2 and CP capabilities, many correctness fixes across decompositions and backends, a much more thoroughly tested codebase, and modernized packaging. Python 3.11 or newer is now required.

Algorithmic improvements

PARAFAC2

• EM-based ALS for PARAFAC2 by @cchatzis in https://github.com/tensorly/tensorly/pull/563
• Hybrid SVD initialization for high rank, with warnings and tests by @MiklosKralik in https://github.com/tensorly/tensorly/pull/617
• Fixes for short slices and random_state handling by @aarmey in https://github.com/tensorly/tensorly/pull/634

CP and randomized CP

• Group Lasso regularization in CP decomposition by @marubekko in https://github.com/tensorly/tensorly/pull/621
• Fix singular matrix error in randomised_parafac and add leverage score sampling by @aarmey in https://github.com/tensorly/tensorly/pull/641
• parafac no longer modifies the init CPTensor or fixed_modes in place by @aarmey in https://github.com/tensorly/tensorly/pull/633
• Shared reconstruction error between CP and NN-CP, fixing masking by @aarmey in https://github.com/tensorly/tensorly/pull/595
• Cap HALS inner NNLS iterations to fix jax/tensorflow test blowups by @aarmey in https://github.com/tensorly/tensorly/pull/626
• Forward random_state in constrained CP random initialization by @AHMETHAKANBEZIR1 in https://github.com/tensorly/tensorly/pull/646

Tucker and tensor train

• Support for factors-only initialization in Tucker by @sebulo in https://github.com/tensorly/tensorly/pull/596
• tucker_to_tensor() now works with partial_tucker() outputs that use modes by @Mithrillion in https://github.com/tensorly/tensorly/pull/593
• TT-cross rejects impossible column ranks before sampling by @AHMETHAKANBEZIR1 in https://github.com/tensorly/tensorly/pull/644
• Update _tt_cross.py by @dontempty in https://github.com/tensorly/tensorly/pull/591

SVD

• random_state added to svd_interface() by @MrBones1102 in https://github.com/tensorly/tensorly/pull/616
• n_eigenvecs capped at min dimension by @aarmey in https://github.com/tensorly/tensorly/pull/634
• Fixed phase multiplier of singular vectors by @Albly in https://github.com/tensorly/tensorly/pull/608
• Fixed sparse SVD backend support by @tandede in https://github.com/tensorly/tensorly/pull/624

Backends and tensor algebra

• The einsum tenalg backend now runs in CI, and the bugs this surfaced are fixed by @aarmey in https://github.com/tensorly/tensorly/pull/625
• Backend inconsistencies fixed and dead backend code removed by @aarmey in https://github.com/tensorly/tensorly/pull/642
• Fixed tenalg backend instance restoration by @AHMETHAKANBEZIR1 in https://github.com/tensorly/tensorly/pull/640
• Clearer choice of modes in tensor contraction by @ax-le in https://github.com/tensorly/tensorly/pull/599
• Batch-axis order preserved in core tensor contraction by @AHMETHAKANBEZIR1 in https://github.com/tensorly/tensorly/pull/638
• Paddle: fixed default device setting and allowed lstsq fallback by @HydrogenSulfate in https://github.com/tensorly/tensorly/pull/583

Packaging and CI

• Packaging moved to pyproject.toml; Python >= 3.11 required by @aarmey in https://github.com/tensorly/tensorly/pull/648
• Python 3.9 dropped, docs build fixed, matplotlib pinned, and CI actions bumped by @aarmey in https://github.com/tensorly/tensorly/pull/627
• Fix flaky tests and docs build warnings by @aarmey in https://github.com/tensorly/tensorly/pull/647

Quality of life improvements and other changes

Documentation, docstrings and many small fixes, plus more reliable tests.

• Fix partial_unfold and add test by @JeanKossaifi in https://github.com/tensorly/tensorly/pull/589
• Escape characters compatible with Python >= 3.12 by @acotino-ignitioncomputing in https://github.com/tensorly/tensorly/pull/600
• Fix soft-sparsity L1 ball projection feasibility and shape by @AHMETHAKANBEZIR1 in https://github.com/tensorly/tensorly/pull/636
• Honour inplace in CPTensor.normalize and return the tensor by @VenishPaneliya in https://github.com/tensorly/tensorly/pull/628
• Fix docstrings and one parameter name that do not match the signatures by @VenishPaneliya in https://github.com/tensorly/tensorly/pull/629
• Document proximal operators by @aarmey in https://github.com/tensorly/tensorly/pull/631
• Document sparse support in partial_unfold by @aarmey in https://github.com/tensorly/tensorly/pull/632
• Add test for non_negative_tucker_hals with algorithm="active_set" by @aarmey in https://github.com/tensorly/tensorly/pull/630

Many thanks to the whole TensorLy team and all contributors, in particular our 13 new contributors: @Mithrillion @ax-le @acotino-ignitioncomputing @cchatzis @Albly @dontempty @tandede @marubekko @sebulo @MiklosKralik @MrBones1102 @VenishPaneliya @AHMETHAKANBEZIR1

Full Changelog: https://github.com/tensorly/tensorly/compare/0.9.0...0.10.0