HP2: Hybrid and precision-guided filter pruning for CNN compression

Abstract

Filter pruning has emerged as a promising approach for compressing Convolutional Neural Network (CNN) models. However, existing methods often lack accuracy in evaluating filter importance and precision in on-demand filter pruning. In this paper, we address these limitations by proposing a novel Hybrid and Precision-guided filter Pruning method (HP²) for CNN compression, driven by two key observations. In particular, our method enhances filter importance evaluation and enables targeted filter pruning, allowing flexible reduction of computational complexity (FLOPs) or memory (parameters). We introduce the Hybrid Importance Score (HIS) to assess precise filter importance by leveraging both filter weights and activations. Moreover, we quantitatively analyze the intricate relationship between FLOPs and parameters, leading to an on-demand pruning strategy that further optimizes FLOPs or parameter reduction. Extensive experiments showcase the superiority of HP² over state-of-the-art CNN compression methods, particularly under high pruning ratios.