Version 1.6.2

Added a reference to the published SLIPMAP paper.

Version 1.6.1

What's Changed

• Faster kmeans when plotting: #15

Version 1.6.0

What's Changed

• Implement Slipmap, a variant based on prototypes that scales better with respect to both time and memory.
• Separate (optional) regularisation functions.
• Use Bayesian optimisation for hyperparameter tuning.
• Remove pre-1.3 deprecations.
• Add optimisation stats to metadata.
• Move paper experiments to frozen branches

Full Changelog: v1.5.2...v1.6.0

Version 1.5.2

Changes:

• Improved sm.plot_dist(scatter=True)
• Improved cluster handling in sm.plot()
• Added example notebook for interactive plots

Version 1.5.1

Changes:

• Fix a isinstance(bool, int) == True oversight

Version 1.5.0

Changes:

• plot: improve labels for matrix plot
• plot: more flexible clusters (not just integers [0..k])
• plot: bars accepts a list of variables
• get_model_clusters: sort the clusters labels based on the embedding
• save and load: add (optional but default) LZMA compression
• metadata.get_variables: fix fallback

Version 1.4.0

Additions:

• Add procedures for hyperparameter optimisation (see slisemap.tuning).
• Added wrappers that combine local_model, local_loss, and coefficients into one name (see, e.g., slisemap.local_models.LogisticRegression).

Changes:

• Change defaults for slisemap.metrics.accuracy (to optimise=False, local_loss=True).
• Add numpy parameter to Slisemap.fit_new.
• Improve the Slisemap.predict function to take more parameters.
• Move entropy to slisemap.metrics.
• Fix some deficiencies in slisemap.utils.Metadata.
• Improve the x-axis labels for the matrix plot.

Version 1.3.1

New features:

• Metadata
  • Add metadata about e.g. variable names and normalisation
  • The metadata is reused for all plots
• More general input
  • Anything that can be turned into a torch.Tensor or numpy.ndarray is accepted
  • This includes dataframe-like objects (also imports variable names)

Improvements:

• Better colourscales for local losses
• Centering before PCA
• Slisemap.load understands map_location from torch.load
• More verbosity levels for optimisation

Changes:

• Remove types from docstrings
• Rename "Fidelity" to "Local loss" in plots

Deprecations:

• With the addition of metadata, some parameters to the plotting functions have been deprecated

Version 1.2.1

What's Changed

• Refactor demo examples + add binder demo to README by @MomoLangenstein in #3
• Point the user towards multiple_linear_regression if they use nd-y by @MomoLangenstein in #5
• Improve docs by @Aggrathon in #6
• Cleanup of properties (some deprecation warnings)
• Update references by @Aggrathon in #7

Full Changelog: v1.1.0...v1.2.1

Version 1.1.0

Main changes in this release:

• Improved documentation.
• Less randomness (only perturb the embedding if a loss of rank is detected).
• New argument only_B in the Slisemap.optimise (the same as Slisemap.lbfgs(only_B)).
• Tweaks to the plots and updated experiments.
• A notebook discussing how to use PyTorch for optimisation.