CHaPR: Efficient Inference of CNNs via Channel Pruning

Abstract

To deploy a CNN on resource-constrained edge platforms, channel pruning techniques promise a significant reduction of implementation costs including memory, computation, and energy consumption without special hardware or software libraries. This paper proposes CHaPR, a novel pruning technique to structurally prune the redundant channels in a trained deep Convolutional Neural Network. CHaPR utilizes a proposed subset selection problem formulation for pruning which it solves using pivoted QR factorization. CHaPR also includes an additional pruning technique for ResNet-like architectures which resolves the issue encountered by some existing channel pruning methods that not all the layers can be pruned. Experimental results on VGG-16 and ResNet-50 models show 4.29X and 2.84X reduction, respectively in computation cost while incurring 2.50% top-1 and 1.40% top-5 accuracy losses. Compared to many existing works, CHaPR performs better when considering an Overall Score metric which accounts for both computation and accuracy.