Filter pruning is one of the most effective ways to accelerate and compress convolutional neural networks (CNNs). In this work, we propose a global filter pruning algorithm called Gate Decorator, which transforms a vanilla CNN module by multiplying its output by the channel-wise scaling factors (i.e.

Especially in [47], a similar two-step framework was proposed, which does not have to train from scratch (Inaccurate description in Page 3 row 83-84). It would be great if authors could describe the new insights of using Taylor expansion to estimate the change of loss along with importance of a filter. It would be great if authors could compare with [26], [47] and ''Data-Driven Sparse Structure Selection for Deep Neural Networks'' (Huang and Wang, 2017) due to the similarity of these ideas. Given the results presented on ImageNet using ResNet-50, which is 0.31 accuracy improvement and 40% FLOPs reduction, it's difficult to compare with (Huang and Wang, 2017) with results of error 26.8 and 66% reduction.

A technically sound approach to pruning networks with experimental evidence of performance, and a clearly written paper.

Filter pruning is one of the most effective ways to accelerate and compress convolutional neural networks (CNNs). In this work, we propose a global filter pruning algorithm called Gate Decorator, which transforms a vanilla CNN module by multiplying its output by the channel-wise scaling factors (i.e. When the scaling factor is set to zero, it is equivalent to removing the corresponding filter. We use Taylor expansion to estimate the change in the loss function caused by setting the scaling factor to zero and use the estimation for the global filter importance ranking. Then we prune the network by removing those unimportant filters. After pruning, we merge all the scaling factors into its original module, so no special operations or structures are introduced.

Filter pruning is one of the most effective ways to accelerate and compress convolutional neural networks (CNNs). In this work, we propose a global filter pruning algorithm called Gate Decorator, which transforms a vanilla CNN module by multiplying its output by the channel-wise scaling factors (i.e. When the scaling factor is set to zero, it is equivalent to removing the corresponding filter. We use Taylor expansion to estimate the change in the loss function caused by setting the scaling factor to zero and use the estimation for the global filter importance ranking. Then we prune the network by removing those unimportant filters.