REPrune: Filter Pruning via Representative Election
This work addresses filter pruning for efficient neural network compression, offering a novel approach that improves upon existing methods but is incremental in the broader context of model compression.
The paper tackles the problem of filter pruning in neural networks by proposing REPrune, a method that selects representative filters via clustering instead of relying on norm-based criteria, achieving over 49% FLOPs reduction with a 0.53% accuracy gain on ResNet-110 for CIFAR-10 and over 41.8% FLOPs reduction with 1.67% Top-1 validation loss improvement on ResNet-18 for ImageNet.
Even though norm-based filter pruning methods are widely accepted, it is questionable whether the "smaller-norm-less-important" criterion is optimal in determining filters to prune. Especially when we can keep only a small fraction of the original filters, it is more crucial to choose the filters that can best represent the whole filters regardless of norm values. Our novel pruning method entitled "REPrune" addresses this problem by selecting representative filters via clustering. By selecting one filter from a cluster of similar filters and avoiding selecting adjacent large filters, REPrune can achieve a better compression rate with similar accuracy. Our method also recovers the accuracy more rapidly and requires a smaller shift of filters during fine-tuning. Empirically, REPrune reduces more than 49% FLOPs, with 0.53% accuracy gain on ResNet-110 for CIFAR-10. Also, REPrune reduces more than 41.8% FLOPs with 1.67% Top-1 validation loss on ResNet-18 for ImageNet.