Dynamic Structure Pruning for Compressing CNNs

Structure pruning is an effective method to compress and accelerate neural networks. While filter channel are preferable other structure methods in terms of realistic acceleration hardware compatibility, with a finer granularity, such as intra-channel pruning, expected be capable yielding more compact computationally efficient Typical utilize static hand-crafted granularity due large search space, which leaves room for improvement their performance. In this work, we introduce novel method, termed dynamic identify optimal granularities pruning. contrast existing methods, the proposed automatically optimizes each layer while training deep To achieve this, propose differentiable group learning designed efficiently learn based on gradient-based groups. The experimental results show that achieves state-of-the-art performance better GPU compared particular, it reduces FLOPs ResNet50 by 71.85% without accuracy degradation ImageNet dataset. Our code available at https://github.com/irishev/DSP.