Rax 0.4.0

What's Changed

• Expose normalize_probabilities as a good normalization for SoftmaxLoss.
• Remove use of initial argument to jax.nn.softmax and jax.nn.log_softmax
• Drop python 3.8 checks and add python 3.11 checks.
• Changes in lambda weights to reduce boilerplate and add new options.
• Fix pytype and clean up types across codebase.
• Minor typo fixes in documentation.

Full Changelog: v0.3.0...v0.4.0

Rax 0.3.0

What's Changed

• New losses: rax.pairwise_soft_zero_one_loss and rax.pairwise_qr_loss.
• Added lambdaweight functionality for all pairwise losses (rax.*_lambdaweight).
• Added segmented data support for all losses and metrics via the segments= keyword argument.
• New example: Segmentation.

Full Changelog: v0.2.0...v0.3.0

Rax 0.2.0

What's Changed

• New losses: rax.poly1_softmax_loss and rax.unique_softmax_loss.
• New example: T5X integration.
• Removed dependencies' version upper-bounds.
• Minor documentation fixes.

Full Changelog: v0.1.0...v0.2.0

Rax 0.1.0

This is the initial release of Rax: a Learning-to-Rank (LTR) library built on top of JAX. It includes the following functionality:

• Ranking losses (rax.*_loss).
• Ranking metrics (rax.*_metric).
• Function transformations (rax.*_t12n).