Convolutional neural network model compression method and device

Abstract

The invention discloses a convolutional neural network model compression method and device, and the method comprises the following steps: S1, carrying out sparse training on an importance factor gamma of each convolution channel of an original convolutional neural network model by employing dynamic regularization constraint, enabling the effective weight of the model to be concentrated in a channel with a relatively large importance factor as much as possible, and obtaining a sparse convolutional neural network model; and S2, according to the size of the channel importance factor of the sparse convolutional neural network model, executing global channel pruning to cut off unimportant channels, and obtaining a compressed pruned convolutional neural network model.

Description

technical field [0001] The invention relates to the technical field of convolutional neural networks, in particular to a compression method and device for a convolutional neural network model. Background technique [0002] The field of machine vision is an important branch of artificial intelligence, and it is widely used in scenarios such as autonomous driving and security. Deep learning algorithms represented by convolutional neural networks perform well in tasks such as object detection, but convolutional neural network models are difficult to deploy on edge computing platforms where memory and computing resources are scarce due to the huge amount of parameters and computing resource overhead. In practical applications, in addition to the high-precision requirements of the algorithm, a large number of application fields also have strict requirements on the running speed of the algorithm. Under the condition of certain computing resources, there is a contradiction between...

AI Technical Summary

Problems solved by technology

[0004] However, the above methods have their own shortcomings: the above pruning algorithms usually do not perform sparse training or perform sparse training with fixed regularization term constraints, which cannot fully exploit the redundancy in the convolutional neural network, resulting in low pruning efficiency and compression. The loss of precision is large; the quantization method reduces the bit width of the convolutional neural network weight storage, but does not reduce the calculation amount of the convolutional neural network; the method of weight decomposition changes the weight structure of the network, and it is difficult to combine with other compression methods to form a combination Compression strategy, limited scalability

Images