CSP: Channel and Space Pruning for Compressing Deep Convolutional Neural Networks

Pruning is widely researched as one of the effective methods for model compression. However, existing pruning methods still have shortcomings in achieving high model compression rates and maintaining model performance. To this end, we propose a general method called Channel And Space Pruning (CSP) for compressing deep convolutional neural networks (DCNNs). The proposed CSP method comprises Coarse-grained pruning and Fine-grained pruning, which operate on the convolutional kernels from both channel and spatial dimensions to achieve model compression. The Coarse-grained pruning utilizes a Bottleneck-based maximum flow mechanism to ensure maximum flow of inputs and outputs in DCNNs, thereby maintaining model performance. Additionally, a Bottleneck-based dynamic updating mechanism is employed to evaluate the importance of convolutional kernels, allowing for the pruning of redundant kernels while maintaining model performance. The Fine-grained pruning further compresses the model based on Coarse-grained pruning. To ensure that the parameters of the convolutional kernels obtained through Coarse-grained pruning are not spatially redundant, Fine-grained pruning employs the Parameters-free spatial attention mechanism (PSAM) for spatial sparsification of the kernels. The proposed CSP method demonstrates promising results compared to state-of-the-art (SOTA) methods in classic DCNN models.