Image model training method, image processing method, chip, equipment and medium

Abstract

The invention discloses an image model training method, an image processing method, a chip, equipment and a medium. The method comprises the steps of obtaining a sample image; wherein one sample imageis associated with one original network output; inputting the sample image into a preset neural network model containing the initial parameters, and performing symmetric quantization on the sample image to obtain a first quantization network output; determining a first cosine similarity between the first quantization network output and the original network output; performing optimization on the preset neural network model according to the first cosine similarity, a preset similarity threshold and a preset mixing precision quantification method, and obtaining preset network output of the preset neural network model after optimization; and determining a loss value between a preset network output and the original network output, and recording a preset neural network model as an image processing model when a convergence condition is preset for the loss value. According to the invention, the parameters of the preset neural network model are reduced, and the calculation rate of the preset neural network model is improved.

Description

Technical field[0001]The invention relates to the field of smart chips, in particular to an image model training, image processing method, chip, equipment and medium.Background technique[0002]With the development of science and technology, artificial intelligence technology is also applied in various scenarios, such as smart robots, smart homes, smart security and other scenarios. In the above scenarios, several integrated smart chips or separate smart chips are often used to implement different functions in each scenario. In different scenarios, the requirements for smart chips are different.[0003]The neural network model parameters in the general smart chip are all float32 type operators, but as the parameters of the neural network model in the smart chip become more and more, when the smart chip wants to obtain higher performance and energy consumption ratio, it has to use low Bit operator calculations, for example, use int4 / int8 / int16 / float16 instead of the original model float3...

AI Technical Summary

Problems solved by technology

[0004] The embodiment of the present invention provides an image model training, image processing method, chip, device and medium to solve the problem that the number of bits quantized by the neural network cannot balance the performance and precision of the smart chip

Images