Deep neural networks (DNNS) have obtained compelling performance among many visual tasks by a significant increase in the computation and memory consumption, which severely impede their applications on resource-constrained systems like smart mobiles or embedded devices. To solve these problems, recent efforts toward compressing DNNS have received increased focus. In this paper, we proposed an effective end-to-end channel pruning approach to compress DNNS. To this end, firstly, we introduce additional auxiliary classifiers to enhance the discriminative power of shallow and intermediate layers. Secondly, we impose L1-regularization on the scaling factors and shifting factors in batch normalization (BN) layer, and adopt the fast and iterative shrinkage-thresholding algorithm (FISTA) to effectively prune the redundant channels. Finally, by forcing selected factors to zero, we can prune the corresponding unimportant channels safely, thus obtaining a compact model. We empirically reveal the prominent performance of our approach with several state-of-the-art DNNS architectures, including VGGNet, and MobileNet, on different datasets. For instance, on cifar10 dataset, the pruned MobileNet achieves 26.9x reduction in model parameters and 3.9x reduction in computational operations with only 0.04"/0 increase of classification error.
