IPBFS: Inference-aware Pruning with Bayesian Optimization based on Filter Similarity

Deep neural networks (DNNs) are significant tools for solving different problems in abundant fields. However, DNNs have large-scale parameters and weight redundancy and require high resources. Therefore, using DNNs is limited to devices such as wearable devices, mobile phones, and other edge devices that do not have enough resources to run. Using neural network pruning techniques, neural network models can be accelerated, and deep neural networks can be deployed on edge devices. However, the presence of filters with similar feature maps in each convolution layer increases redundancy, parameters, and the count of floating-point operations (FLOP) of deep models. In this paper, by combining the filter similarity algorithm and Bayesian optimization, the filters with a high percentage of similarity between their output feature maps and have a trivial effect on the model accuracy are automatically pruned. The proposed method has been applied to VGG16, ResNet20, and ResNet39 models on CIFAR10 and CIFAR100. Based on the best results, the FLOP and the count of parameters of the VGG16-CIFAR10 have decreased by 92.53% and 54.98%, and in the VGG16-CIFAR100 by 92.26% and 55.3%, respectively. Also, in the ResNet20 model, the count of parameters and FLOP operations have decreased by 31.48% and 13.89%, and in the ResNet39 model by 41.07% and 17.87%.