Deep neural network compression method

Abstract

The invention relates to a deep neural network compression method, which comprises the following steps of: setting the value of a P parameter by using at least one weight of a deep neural network; grouping the weights in a state that every P weights form one group; branch pruning retraining is executed, so that branches are uniformly grouped in a state that only one weight of each group is non-zero and the other weights are zero. Therefore, the compression rate of the deep neural network is adjusted, and the reduction rate of the deep neural network is adjusted.

Description

technical field [0001] A method of compression, in particular of a deep neural network with uniform branching groupings. Background technique [0002] Deep learning has the highest accuracy in image, speech recognition, autonomous driving, machine translation, etc. Almost all successful deep learning techniques are built on top of neural networks. DNN is the simplest neural network, and existing technologies are usually connected by CNN, RNN, LSTM and other neural networks. After the DNN is imported into the embedded platform, it has a wide application impact on mobile computing. However, the size of DNN models is usually very large, sometimes close to 100 million bytes, which limits the improvement of large-scale models on IoT platforms and edge devices. For example, mobile computing platforms are limited in CPU speed, memory, and battery life. So, to apply these neural networks to these edge devices. Although the prior art can reduce parameters and reduce memory usage...

AI Technical Summary

Problems solved by technology

However, the size of DNN models is usually very large, sometimes approaching 100 million bytes, which limits the improvement of large-scale models on IoT platforms and edge devices

For example, mobile computing platforms are limited in CPU speed, memory, and battery life

Images