CpdsConv: Continuous Pruning for Depthwise Separable Convolution

Current structured pruning methods typically employ fixed thresholds and rank channels based on evaluation metrics for pruning, leading to significant drops in network accuracy post-pruning. This paper proposes a novel continuous pruning method for depthwise separable convolution, named CpdsConv, to address this issue. CpdsConv uses a two-stage training approach combining depthwise separable convolutions (TSConv) and learnable threshold sparse convolutions (LTSConv) to prune the network, aiming to enhance accuracy while compressing the model. To validate the effectiveness of the proposed method, we conducted experiments on the VGG-16, MobileNet, GoogLeNet, and ResNet-56 models. The results demonstrated that, while maintaining the original accuracy, the proposed method significantly reduced the number of parameters and computational burden in these models, showing a better balance overall. Notably, in the ResNet-56 experiment, the parameters were reduced by 37.9%, with a simultaneous accuracy improvement of 0.81%. This indicates that our method not only effectively reduces computational complexity but also significantly enhances model performance.