FLOPs-efficient filter pruning via transfer scale for neural network acceleration

Abstract

Model pruning is a useful technique to reduce the computational cost of convolutional neural networks. In this paper, we first propose a simple but effective filter level pruning criterion, which assesses the importance of a filter by exploring the transfer scale (TS) of its feature maps in the next layer. The principle is that for a trained CNN model, an important filter should have strong connections with the next layer, otherwise the transfer scale of its feature map will be low and hence removing it will have little influence on the network. Besides, we observe that filters from the computationally-intensive layers are more sensitive to pruning, which makes it difficult to further compress the floating-point operations (FLOPs) of the model without reducing accuracy. To solve this problem, we propose a FLOPs-efficient group Lasso approach for TS to guide the network to use fewer filters in the computationally-intensive layers, which leads to better FLOPs compression performance after pruning. We refer to the proposed method as FETS. Compared with the state-of-the-art methods, our FETS achieves similar or better accuracy, but with significantly larger FLOPs compression ratio. In particular, with VGG-16, ResNet-56 and DenseNet-40 on CIFAR-10, we achieve similar or better accuracies than other methods, with only 48%, 64% and 58% of the FLOPs. With ResNet-50 on ImageNet, we also achieve a relative FLOPs reduction of 30%.