PEFT Test Release

Sorry we have not been keeping up on these release
Please see the ChangeLog

Minor Release 0.55

weightwatcher 0.5.5 has been released to pypi
pip install weightwatcher
This minor release includes several new, experimental features, including

• rand_distance metric, a new method for measuring how well a layer is trained
• fix_fingers='xmin_peak' , which fit a the entire ESD, not just the long tail, to a power law
• fix_fingers='clip_xmax' , which attempts to correct for spuriously large values of alpha
• SVDSharpness Transform, which clips off large elements of W using RMT
• SVDSmoothness Transform, which forms a low rank approximation of the model
• StackedLayerIterator, which can be used for very small models
• Save Figure now saves all images to a subfolder
• Vector Entropy and Localization Metrics now plotted

Minor Release 0.4.8

ONNX support added
IntraLayer correlations added (experimental)

Minor release 0.4.6

Bug fix to 0.4.5

0.45 minor release

added:

• custom logging
• matrix entropy
• matrix rank
• savefig now saves every layer
• savefig added
• color coding on ESD plots change
• RMTUtil factored a bit

Bug fix for 0.4.1

0.4.1 was missing RMTUtil and not tested properly in a sandboxed environment
Please upgrade to 0.4.2

minor release 0.4.1

Fixed minor bug with ww2x=True and min_evals

WeightWatcher version 0.1.2 Analyze Weight Matrices of Deep Neural Networks

Release 0.1.2 Added

support for pandas

• get_summary(pandas=True)
• get_details()

SoftRank and SpectralNorm metrics

Support for Huggingface BERT models

WeightWatcher v0.1.1 to analyze weight matrices of Deep Neural Networks

First release of WeightWatcher