Optimization Based Layer-Wise Pruning Threshold Method for Accelerating Convolutional Neural Networks

Among various network compression methods, pruning has developed rapidly due to its superior performance. However, the trivial threshold limits performance of pruning. Most conventional methods are based on well-known hard or soft techniques that rely time-consuming handcrafted tests domain experience. To mitigate these issues, we propose a simple yet effective general method from an optimization point view. Specifically, problem is formulated as constrained program minimizes size each layer. More importantly, our together with works achieves better across scenarios many advanced benchmarks. Notably, for L1-norm algorithm VGG-16, higher FLOPs reductions without utilizing sensibility analysis. The ratio boosts 34% 53%, which huge improvement. Similar experiments ResNet-56 reveal that, even compact networks, competitive skipping any sensitive layers.