Face recognition model compression method, medium and terminal based on model distillation

Abstract

The present invention provides a face recognition model compression method, medium and terminal based on model distillation. The algorithm includes: obtaining an initial network model for face recognition; performing channel pruning processing on the initial network model to obtain channel pruning The processed model is fine-tuned to the model after the pruning process; the feature layer of the initial network model is used as the weight, and the feature layer of the initial network model is fitted with the fine-tuned model feature layer to complete the distillation of the face recognition model ; The face recognition model compression method based on model distillation in the present invention does not need the final classification layer, which greatly improves the training speed and saves video memory. Combined with model pruning, by removing unimportant weights, the model is compressed and the speed is improved. It can refine the knowledge of large networks into small networks, improve the accuracy of face recognition, and meet the needs of large-scale classification tasks such as face recognition.

Description

technical field [0001] The invention relates to the field of computer applications, in particular to a face recognition model compression method, medium and terminal based on model distillation. Background technique [0002] In recent years, Deep Neural Network (DNN, Deep Neural Network), as a very effective problem-solving method, has been widely used in many fields such as computer vision and speech recognition, such as convolutional neural network CNN (Convolutional neural network) in computer Many traditional visual problems (classification, detection, segmentation) have surpassed traditional methods. When using deep networks to solve problems, they usually use complex networks to collect more data in order to obtain better performance. However, what followed was a sharp increase in the complexity of the model. The intuitive performance is that there are more and more model parameters, the model is getting bigger and bigger, and the required hardware resources (memory, v...

AI Technical Summary

Problems solved by technology

[0004] However, for the face recognition task, due to the large amount of parameters and the existence of redundancy

Although there are many current distillation methods, they are not suitable for face recognition tasks, because the last classification layer of face recognition has a large number of categories and a huge amount of parameters, while traditional knowledge distillation outputs the probability of each category through the teacher model, and then Use a small model to fit this distribution, but for face recognition up to millions of classes, this distribution is difficult to fit

At present, the existing model distillation algorithm has little effect on large-scale classification tasks such as face recognition, and cannot meet the actual needs of its large number of parameters.

Images