Start and Stop functionality as well as some plot-bugfixes

• Plots no longer shrink with each epoch when recording layer saturation
• the torchcallback now features a stop() and resume() function which stop and resume the recording of stats and aggregation of saturation values.

Bugfix covariance computation

0.1.39

Update version.py

removed a rogue print statement

0.1.38

Update version.py

feature map downsampling

It is now possible to downsample feature maps to a minimum size in order to keep computation costs a low as possible

resume training is now possible

It is now possible to resume the training by setting the initial_epoch parameter.

Performance Improvements

Tripled the performance by removing redundant covariance computation

Refactoring, Result Reconstruction

• It is now possible to provide a list of writers or corresponding string keys for the "save_to" parameter
• It is now possible to save the covariance matrix, however only npy-writer supports saving the covariance matrix
• Added a utility function that allows reconstructing result csvs with saturation and intrinsic dimensionality on arbitrary thresholds. Works only of npy save strategy is used for the run.
• Adjusted look and configuration of all plots
• Intrinsic Dimensionality can be computed
• Cleaned differen computation strategies for intrinsic dimensionality, such that all work exactly the same now.
• Saturation computation is now fully implemented in double precision in order to avoid rounding errors

PCA Layers and TorchCovariance Bugfixes

Bugfix on Covariance

Added PCA Layers

Bugfix for layer saturation + wip implementation of PCA Layers

1.31.0

fixed bugs with pooling strategies. Enabled channelwise-pooling a new…

Added channelwise saturation computation

Added channelwise saturation computation, which should be more accurate than mean-saturation