NFP: A No Fine-tuning Pruning Approach for Convolutional Neural Network Compression

Pruning for Convolutional neural network has proved to be an effective approach to reduce memory and computation. In this paper, we propose a novel pruning approach called NFP(No Fine-tuning Pruning), which calculates the contribution of the pruned convolutional filters to the next convolutional layer and compensates the contribution to the bias of the next convolutional filters. Due to the compensation, the accuracy of pruned network will be almost same as the original network. Therefore, the NFP method can only take a few moment to obtain an accuracy retaining compact network without fine-tuning. We demonstrate the effectiveness of our approach on some CNN models. For ResNet-50, NFP can reduce the number of parameters by 34% and the FLOPs by 37% without loss of accuracy. For object detection task, our approach also achieves an excellent result in YOLO-V3.