Convolution Kernel Pruning Algorithm Based on Average Percentage of Zeros and Data Distribution Similarity

Pruning convolutional kernels is a crucial method for achieving model lightweighting. However, current APoZ (Average Percentage of Zeros) based pruning algorithms often overlook the similarity in data distribution among convolutional kernels, resulting in significant degradation in accuracy after pruning. To address this issue, we propose a novel convolutional kernel pruning algorithm based on APoZ_S (Average Percentage of Zeros and Similarity) values. The APoZ_S criterion integrates APoZ and similarity for convolutional kernel pruning. We utilized Gaussian mixture model (GMM) to model the weight data of the convolution kernel and calculated the similarity fraction of the convolution kernel according to the modeling results. We then added the APoZ value to obtain the APoZ_S value. Finally, we performed convolution kernel pruning by setting a threshold value. Test results from different network models on the cifar-10 dataset demonstrate that compared with the traditional APoZ-based convolutional kernel pruning algorithm, our proposed algorithm yields higher accuracy in output results under different pruning rates.