v1.5.1

• Bump dependencies

Full Changelog: v1.5.0...v1.5.1

v1.5.0

• Neural network enhancement by changing the input to the PCA space and switching to RMSNorm.
• Gaussian Mixture and consensus K-Means initialization
• Improved OOM data handling
• Multi-GPU training and inference

Note: these releases introduce breaking changes which might make models trained on prior versions not to work as expected. Please use a version previous to 1.5.0 if that's the case.
Full Changelog: v1.4.1...v1.5.0

Update dependencies

• Update dependencies to avoid breaking changes and speedup data reading.
• Drop columns containing only NaNs from the input matrix.

Mean imputation

• Added mean imputation for missing data
• Mean imputation is now the default option for imputing

Full Changelog: v1.3.1...v1.4.0

Bugfix: warmup training

• Fixes #22 (warmup training breaking for different initializations).
• Removed unnecessary print in training loop.

Improved stability

• Added warmup training for initialization of Q to improve training stability.
• Changed default P initialization to 'pckmeans'.
• Changed convergence check so it is performed after 15 epochs.
• Changed default learning rate to 1e-5 instead of 1e-4.
• Added acceleration for Apple Metal in compatible devices (M1 and M2 family).

Stable release

• Update docs

Improved version handling

• Automatic version bump
• Changed version log in entrypoint

Improved file extension handling. Dependency fix.

• Fixed bug where filenames containing . before the actual extensions caused the SNPReader class to crash.
• Pin scikit-learn to avoid an error in dask-ml's IncrementalPCA

Dask backend

Switched to a dask backend to handle new arrays.
New packaging format.
Improved paths handling.
Updated CI workflows.