Neural network accelerator model conversion method and device

Abstract

The invention discloses a neural network accelerator model conversion method and device. The method comprises the steps of obtaining a to-be-converted neural network model; analyzing the model network structure file to obtain all network layers of the model; reconstructing the network layer; mapping to operator nodes supported by a neural network accelerator; and finally, serializing the converted operator nodes and model weights according to a network topology structure to generate a target file. The device comprises a neural network model construction module, a reconstruction module, a mapping module and a serialization module. According to the invention, the problem of multi-adaptation difficulty of deployment of various format models on neural network accelerator equipment is solved, model conversion can be efficiently carried out, and the model format adaptive to the neural network accelerator is generated.

Description

technical field [0001] The invention relates to the field of artificial intelligence, in particular to a neural network accelerator model conversion method and device. Background technique [0002] The current neural network technology is developing rapidly, and various machine learning frameworks have emerged as the times require. Various machine learning frameworks have their own advantages and disadvantages. Therefore, in academia and industry, various machine frameworks coexist, which has also led to differences. The models trained by the framework have different file formats. For ASIC (Application Specific Integrated Circuit, Application Specific Integrated Circuit) devices such as neural network accelerators, various model file formats need to be adapted, which greatly hinders the neural network model from being used in ASIC devices. deployment efficiency. [0003] Existing model format files must be read based on the framework or protocol corresponding to the format....

AI Technical Summary

Problems solved by technology

[0002] The current neural network technology is developing rapidly, and various machine learning frameworks have emerged as the times require. Various machine learning frameworks have their own advantages and disadvantages. Therefore, in academia and industry, various machine frameworks coexist, which has also led to differences. The models trained by the framework have different file formats. For ASIC (Application Specific Integrated Circuit, Application Specific Integrated Circuit) devices such as neural network accelerators, various model file formats need to be adapted, which greatly hinders the neural network model from being used in ASIC devices. deployment efficiency

[0003] Existing model format files must be read based on the framework or protocol corresponding to the format. For example, the TensorFlow model file uses Protocol Buffers. If you want to quickly read and obtain the network layer information in the model file, you must use TensorFlow Software packages are required, which is very cumbersome on the hardware; at the same time, the model format saved by most machine learning frameworks contains a lot of information that is useless for the neural network accelerator to run the model (such as the learning rate required for training, batch data size, etc. ), which are useless information in neural network accelerators (mainly used for model inference)

In short, the traditional neural network accelerator must use the machine learning framework program to use its saved model files, and the use of the framework for the neural network accelerator will have the problem of adapting different frameworks, which has great limitations

Images