Data Agnostic Filter Gating for Efficient Deep Networks

Filter pruning is essential for deploying a well-trained CNN model on edge computation devices with a target computation budget (\eg, FLOPs). Current filter pruning methods mainly focus on leveraging feature maps to analyze the importance of filters, and prune those with less impact on the value of the CNN?s loss function, thereby ignoring the variance of input batches to differences in sparse structure over the filters. This paper proposes a data-agnostic filter pruning method, which uses an auxiliary network called the Dagger module to prune with pre-trained weights as input. This makes our results not only related to batch information, but also related to training history. Experimental results on CIFAR-10 and ImageNet datasets show that the proposed filter pruning method surpasses the state-of-the-art.