Release Release v1.0 · TimDettmers/sparse_learning

Release v1.0: Bug fixes, New ImageNet Baselines

Changed to boolean indexing for PyTorch 1.2 compatibility.
Fixed an error where an error can occur for global pruning algorithms if very few weights were removed for a layer.
Removed momentum reset. This feature did not have any effect on performance and made the algorithm more complex.
Fixed an error where two layers of VGG16 were removed by use of the remove_weight_partial_name() function. Results were slightly degraded, but weights needed for dense performance and relative ordering compared to other methods remained the same.

Evaluation script can now aggregate log files organized in a folder hierarchy. For each folder results will be aggregated.
Added decay schedule argument. One can choose between Linear and Cosine prune rate decay schedules.
Added new ImageNet baseline which is based on the codebase of Mostafa & Wang, 2019.
Added a max-thread argument which can be used to set the total maximum data loader threads for training, validation and test set data loaders.