Empirical evaluation of filter pruning methods for acceleration of convolutional neural network

Training and inference of deep convolutional neural networks are usually slow due to the depth of the network and the number of parameters in the network. Although high-performance processors usually accelerate the training of these networks, their use on resource-constrained devices is still limited. Several compression-based acceleration methods have been presented to optimize the performance of neural networks. However, their use and adaptation are still limited due to their adverse effects on the network structure. Therefore, different filter pruning methods have been proposed to keep the network structure intact. To better solve the above limitations, we first propose a detailed classification of model acceleration method to explain the different ways of enhancing the inference performance of the convolutional neural network. Second, we present a broad classification of filter pruning methods including the comparison of these methods. Third, we present an empirical evaluation of four filter pruning methods to understand the effects of filter pruning on model accuracy and parameter reduction. Fourth, we perform several experiments with ResNet20, a pre-trained CNN, and with the proposed custom CNN to show the effect of filter pruning on them. ResNet20 is used to address the multiclass classification using CIFAR 10 dataset and custom CNN is used to address the binary classification using Leukaemia image classification dataset that includes low-information medical images. The experimental results show that among the four filter pruning methods, the soft filter pruning method best preserves the accuracy of the original model for both ResNet20 and the custom CNN. In addition, the sampling-based filter pruning method shows the highest reduction of 99.8% in parameters on custom CNN. The overall results show a reasonable pruning ratio within five training epochs for both the pre-trained CNN and custom CNN. In addition, our results show that pruning redundant filters significantly reduces the model size, and number of floating point operations.