To Filter Prune, or to Layer Prune, That Is the Question

Recent advances in pruning of neural networks have made it possible to remove a large number filters or weights without any perceptible drop accuracy. The parameters and that FLOPs are usually the reported metrics measure quality pruned models. However, gain speed for these models is often overlooked literature due complex nature latency measurements. In this paper, we show limitation filter methods terms reduction propose LayerPrune framework. presents set layer based on different criteria achieve higher than similar advantage over result fact former not constrained by original model’s depth thus allows larger range reduction. For each method examined, use same importance criterion calculate per-layer score one-shot. We then prune least important layers fine-tune shallower model which obtains comparable better accuracy its filter-based counterpart. This one-shot process from single path like VGG before fine-tuning, unlike iterative pruning, minimum per required allow data flow constraint search space. To best our knowledge, first examine effect metric instead multiple networks, datasets hardware targets. also outperforms handcrafted architectures such as Shufflenet, MobileNet, MNASNet ResNet18 7.3%, 4.6%, 2.8% 0.5% respectively budget ImageNet dataset (Code available at https://github.com/selkerdawy/filter-vs-layer-pruning).