Weight precision configuration method, weight precision configuration device, weight precision configuration equipment and storage medium

Abstract

The embodiment of the invention discloses a weight precision configuration method and device, equipment and a storage medium. The method comprises the following steps: determining a current recognition rate threshold value from at least two candidate recognition rate threshold values smaller than the target recognition rate threshold value, reducing and adjusting the weight precision correspondingto each layer in the neural network based on the current recognition rate threshold; and training the reduced and adjusted neural network to adjust the weight parameter value of each layer, the training target being to improve the recognition rate of the reduced and adjusted neural network, and determining the final configuration result of the weight precision of each layer according to the relationship between the current recognition rate and the target recognition rate threshold. By adopting the technical scheme, the resource utilization rate in the artificial intelligence chip bearing theneural network can be improved, the chip performance can be improved and the chip power consumption can be reduced under the condition of ensuring the identification rate of the neural network.

Description

technical field [0001] The embodiments of the present invention relate to the technical field of artificial intelligence, and in particular, to a weight precision configuration method, device, device, and storage medium. Background technique [0002] With the vigorous development of big data information networks and smart mobile devices, a large amount of unstructured information has been generated, accompanied by a sharp increase in the demand for high-efficiency processing of these information. In recent years, deep learning technology has developed rapidly, and has achieved high accuracy in many fields such as image recognition, speech recognition, and natural language processing. However, the vast majority of deep learning research today is still based on traditional von Neumann computers. Due to the separation of processors and memory, von Neumann computers not only consume high energy and have low efficiency when dealing with large and complex problems, but also The n...

AI Technical Summary

Problems solved by technology

[0004] At present, deep learning algorithms can work with different data precision, and high precision can achieve better performance (such as accuracy or recognition rate), but after being applied to artificial intelligence chips, the storage cost and calculation cost are relatively high, while Low precision can trade some degree of performance loss for significant savings in storage and computation, making the chip highly power-efficient

In the current common AI chips, due to the different requirements for calculation precision, the processing chip also needs to provide storage support for multiple data precision, including integer (integer, Int) and floating-point (floating-point, FP), etc., such as 8 Bit integer (Int8), 16-bit floating point (FP16), 32-bit floating point (FP32) and 64-bit floating point (FP), etc., but the weight precision of each layer of the neural network carried in the brain-like chip is the same , which makes the weight accuracy configuration scheme in artificial intelligence chips not flexible enough and needs to be improved

Images