Package v0.0.11

New features

• The new version of PiecewiseLinearEmbeddings used in the TabM project was added (in the TabM paper, it is only briefly mentioned in the appendix). It is now recommended by default. To use it, pass the version argument as follows: PiecewiseLinearEmbeddings(..., version="B").
• PiecewiseLinearEmbeddings and PiecewiseLinearEncoding are now significantly faster (due to the new, Python-loop-free implementation).

Documentation

Overall, the documentation was significantly improved. In particular:

• The practical recommendations were revisited and updated.
• For some embedding modules, hyperparameter-related recommendations were revisited. In particular, that applies to the periodic embeddings.
• The usage sections of all embedding modules were improved.

Package v0.0.10

New features

• The package now directly provides LinearEmbeddings instead of importing this class from another package.

Performance

• Significantly improve efficiency of PeriodicEmbeddings and PiecewiseLinearEmbeddings for cases when there is more than one object in a batch (see #19 for details). The larger the input dimensions -- the larger the speedup.
• Improve efficiency of LinearEmbeddings.

Changed

• Remove the dependency on rtdl_revisiting_models.