Neural network pruning consists of reducing the size of a network by removing parameters. In this work, we introduce ShrinkBench, an open-source library to facilitate standardized evaluation of neural network pruning methods. ShrinkBench simplifies using standardized datasets, pretrained models, and evaluation metrics for implementing pruning methods. In addition to describing the functionality of ShrinkBench, we demonstrate its utility by using it to implement and evaluate several pruning methods. We show that ShrinkBench’s comprehensive evaluation can prevent common pitfalls when comparing neural network pruning methods.
José Javier González Ortiz
