A Multi-Scale Automatic Progressive Pruning Algorithm Based on Deep Neural Network

With the advancement of deep learning, an increasing number of network models have been proposed. To enhance model performance, the structure has become progressively intricate, posing challenges for deployment on resour-ceconstrained devices due to its extensive parameterization and computational demands. Model pruning is a prevalent approach to address this issue. However, most existing pruning methods primarily focus on reducing the model width, and the pruning rate for each layer is often determined based on empirical knowledge. These methods face challenges in achieving a balance between compression ratio and model accuracy after pruning, while also requiring significant time and effort to determine the appropriate pruning rates. In this paper, we propose an automatic multi-scale progressive pruning algorithm. We systematically pruned the model by adjusting both the model width and filter shape, while also incorporating translation invariance into VGGNet through the replacement of traditional pooling with BlurPool. This enhancement aimed to improve both the robustness and accuracy of our model. To evaluate its performance, we conducted experiments on three datasets (CIFAR-10, CIFAR-100, and Mini-Imagenet) using both VGGNet and ResNet architectures. After pruning, VGG16 exhibited a reduction of 91.24% in parameter count and 70.74% in computational load on the CIFAR-10 dataset. similarly, after pruning ResNet56 experienced a decrease of 55.81% in parameter count and 59.63% in computational load without any significant loss in accuracy.