Learning-based Tensor Decomposition with Adaptive Rank Penalty for CNNs Compression

Abstract

Low-rank tensor decomposition is a widely-used strategy to compress convolutional neural networks (CNNs). Existing learning-based decomposition methods encourage low-rank filter weights via regularizer of filters’ pair-wise force or nuclear norm during training. However, these methods can not obtain the satisfactory low-rank structure. We propose a new method with an adaptive rank penalty to learn more compact CNNs. Specifically, we transform rank constraint into a differentiable one and impose its adaptive violation-aware penalty on filters. Moreover, this paper is the first work to integrate the learning-based decomposition and group decomposition to make a better trade-off, especially for the tough task of compression of 1×1 convolution.The obtained low-rank model can be easily decomposed while nearly keeping the full accuracy without additional fine-tuning process. The effectiveness is verified by compression experiments of VGG and ResNet on CIFAR-10 and ILSVRC-2012. Our method can reduce about 65% parameters of ResNet-110 with 0.04% Top-1 accuracy drop on CIFAR-10, and reduce about 60% parameters of ResNet-50 with 0.57% Top-1 accuracy drop on ILSVRC-2012.