0.7.2

bump torchdtw and hooks

0.7.1

Support for optional FASTABX_OUTPUT environment variable.
Set FASTABX_OUTPUT=jsonl to format the CLI output to JSON(L)

0.7.0

Bump to PyTorch 2.10 and use its stable ABI

• The DTW implementation has moved to the torchdtw package, which uses PyTorch's new stable ABI
• Going forward, only the minimum PyTorch version is a requirement: you no longer need to pin fastabx to a specific PyTorch version.

0.6.1

free disk space in github runner

0.6.0

Bump to PyTorch 2.9

Default supported PyTorch version is now 2.9.
If you use a different version, download one of the wheels below.

Notice:
We will start using PyTorch Stable ABI for the DTW extension starting in version 2.10.
This means that we will be able to build a single wheel compatible with every PyTorch variant >= 2.10 for a given platform, but the minimum supported version will therefore become version 2.10.

0.5.5

Fix usage of to_decimal with polars 1.34

0.5.4

Internal changes:

• To build the Dataset, files are now identified by their relative path to the root directory, not just by the filename
• Removed the read_item function for more flexibility in reading .item files
• Add a utility script to unpack concatenated fairseq features
• Use setuptools-scm to extract package version from Git
• Fix build and release workflow

0.5.3

fix reading of item files for polars 1.33.0

0.5.2

• Add the constraints parameter to Score which allows for fine-grain filtering of the triplets inside each cell.
• See https://docs.cognitive-ml.fr/fastabx/examples/constraints.html for more details.

(Version 0.5.1 was removed as there was a bug in the masking when the task was symmetric)

0.5.0

• The default PyTorch version is now 2.8.0. Wheels built with other versions are available in the Releases page.
• Now checks for empty features when building the dataset (681b49f)