Efficient Block Pruning Based on Kernel and Feature Stablization

Abstract

With the development of computer vision research, the architecture of convolutional neural network becomes more and more complex to reach the state-of-the-art performance. Is the complexity of the model necessarily proportional to its accuracy? To answer this, the compression of the network has attracted much attention in the academy and industry. Existing network pruning methods mostly rely on the scoring mechanism of complexity or diversity of kernels to compress the network, and then build the network model after removing the kernels by tuning or training on the input data. These methods are cumbersome and depend on a well-trained pre-trained model. In this paper, we propose an end-to-end block pruning method based on kernel and feature stability by pruning blocks efficiently. To accomplish this, we firstly introduce a mask to scale the output of the blocks, and the L1 regularization term to monitor the mask update. Second, we introduce the Center Loss to guarantee that the feature does not deviate greatly during learning. To converge fast, we introduce fast iterative shrinkage-thresholding algorithm (FISTA) to optimize the mask, by which a more fast and reliable pruning process is achieved. We implement experiments on different datasets, including CIFAR-10 and ImageNet ILSVRC2012. All the experiments have achieved the state-of-the-art accuracy.