Efficient structured pruning based on deep feature stabilization

Sheng Xu, Hanlin Chen, Xuan Gong, Kexin Liu, Jinhu Lü, Baochang Zhang

Abstract: The application of convolutional neural networks (CNNs) in computer vision highly depends on the consumption of computation and memory resources, which affects its development on resource-limited devices. Accordingly, CNN compression has attracted increasing attention. In this paper, we propose an efficient end-to-end pruning method based on feature stabilization (EPFS), which is feasible to be implemented for structured pruning such as filter pruning and block pruning. For block pruning, we introduce a mask to scale the output of structures and the $$\ell _1$$ ℓ 1 -regularization term to sparsify the mask. For filter pruning, a novel $$\ell _2$$ ℓ 2 -regularization term is proposed to constraint the mask along with the $$\ell _1$$ ℓ 1 -regularization. Besides, we introduce the Center Loss to stabilize the deep feature and fast iterative shrinkage-thresholding algorithm (FISTA) to accelerate the convergence of mask. Extensive experiments demonstrate the superiority of our EPFS. On CIFAR-10, EPFS saves $$47.5\%$$ 47.5 % FLOPs on VGGNet with $$1.17\%$$ 1.17 % Top-1 accuracy increase. Furthermore, on ImageNet ILSVRC2012, EPFS reduces $$55.2\%$$ 55.2 % FLOPs on ResNet-18 with o.nly $$1.63\%$$ 1.63 % Top-1 accuracy decrease, which promotes the state-of-the-arts.