Compression method and compression device of neural network model, storage medium and equipment

Abstract

The invention discloses a convolutional neural network model compression method based on transfer learning. The compression method comprises the steps of obtaining a pre-trained transfer learning model; performing compression processing on each convolution layer of the transfer learning model according to a preset compression ratio to obtain a first target network model; performing transfer learning on the first target network model by using the target image data set to obtain a first compressed target model; according to a predetermined rule, selecting a part of convolution layers in the first compression target model for compression processing to obtain a second target network model; and performing transfer learning on the second target network model by using the target image data set toobtain a second compressed target model. Advantages of transfer learning and convolution compression are complemented, on the premise that high accuracy is guaranteed, model complexity is reduced, operation speed is increased, compression processing is performed on part of convolution layers, the model complexity can be further reduced, and sudden drop of the model accuracy is avoided.

Description

Technical field [0001] The present invention belongs to the field of information technology, and in particular, relates to a compression method and a compression device for a convolutional neural network model based on transfer learning, a computer-readable storage medium, and computer equipment. Background technique [0002] The basic principle of transfer learning is to adapt a model trained on a problem to a new problem through simple adjustments. Due to the complexity of large-scale networks, training a complex convolutional neural network requires a lot of labeled data, and it takes a long time. Transfer learning is a workaround to solve the problem of large amounts of labeled data and training time. In the case of sufficient data, the effect of transfer learning is not as good as complete retraining. But the training time and the number of training samples required for migration learning are much less than training a complete model. And you can get good accuracy. [0003]...

AI Technical Summary

Problems solved by technology

However, the disadvantage of the existing transfer learning method is that it can only provide a model with the same computational complexity as the neural network trained by the source data.

[0004] The compression algorithm of the convolutional neural network model is a common algorithm to reduce the complexity of the model, but the current compression algorithm has two shortcomings. One is that when the target data is insufficient, the correct rate of the trained model is low. Compression will cause some convolutional layers to be over-compressed, and even some convolutional layers will be completely compressed, which will cause the network model to be incomplete and the accuracy rate will drop sharply

Images