Resumen
As the number of layers of deep learning models increases, the number of parameters and computation increases, making it difficult to deploy on edge devices. Pruning has the potential to significantly reduce the number of parameters and computations in a deep learning model. Existing pruning methods frequently require a specific distribution of network parameters to achieve good results when measuring filter importance. As a result, a feature map similarity score-based pruning method is proposed. We calculate the similarity score of each feature map to measure the importance of the filter and guide filter pruning using the similarity between the filter output feature maps to measure the redundancy of the corresponding filter. Pruning experiments on ResNet-56 and ResNet-110 networks on Cifar-10 datasets can compress the model by more than 70% while maintaining a higher compression ratio and accuracy than traditional methods.