Semantic-Driven Automatic Filter Pruning for Neural Networks

Yang Guo, Wei Gao

Abstract: Filter pruning is widely used for neural network compression. However, existing methods mostly judge importance of filters by measuring magnitudes or distributions of data in weight parameters or feature maps, but neglect semantics of filters, which weakens interpretability. Meanwhile, layer-wise compression ratios are assigned automatically only under given FLOPs by architecture search algorithms or just manually. It takes a lot of time and is impractical. In this paper, a novel semantic-driven automatic filter pruning method is proposed to solve the above problems. Filters are treated as semantic detectors, and the semantic-driven importance criteria is developed by evaluating correlations between input tasks and feature maps. Besides, inspired by the phenomenon that each filter has deterministic impact on FLOPs and network parameters, we provide a new efficient adaptive compression ratio allocation strategy based on differentiable pruning approximation by considering FLOPs and parameters simultaneously. Our method is validated with extensive experiments on various neural networks.