Pruning method based on discrete cosine transform channel importance score

Abstract

The invention discloses a discrete cosine transform channel importance score-based pruning method, which comprises the following steps of: transforming an output feature map in a deep convolutional network by using discrete cosine transform, and calculating a corresponding channel importance score to determine the importance degree of a channel, then, according to a set pruning rate, preferentially subtracting channels with relatively low channel importance scores so as to obtain a target network, and finally, initializing the target network by the reserved channels to carry out adjustment and optimization so as to reduce model parameters, calculation amount and the like under the condition of keeping precision. According to the method, the effectiveness of channel pruning is greatly improved, the method is extremely simple and can be flexibly applied to different application fields, and a reliable and efficient method is provided for neural network model compression.

Description

technical field [0001] The present invention relates to the field of artificial intelligence model compression, and more specifically, to a pruning method based on discrete cosine transform channel importance scores. Background technique [0002] Deep convolutional neural networks have achieved success in many fields such as image classification, object detection, semantic segmentation, text detection, video understanding, and super-resolution. However, the improvement of its performance is often accompanied by the cost of huge storage and computing resources. How to reduce the amount of parameters and calculations of the network model while maintaining high accuracy is the current challenge in the deployment of deep convolutional neural networks. one. In order to solve this problem, researchers have proposed model compression methods such as quantization, channel pruning, knowledge distillation, and tensor decomposition to solve the problem of model compression in a comple...

AI Technical Summary

Problems solved by technology

[0003] The pruning method in the prior art requires a complex model and a huge amount of computational support, and it also needs to transform the input image into the frequency domain for pruning. Not only is the calculation complicated, but also the existing channel pruning method The evaluation of channel importance is limited, thus limiting the efficient compression of the model

Images