Pruning threshold determination method and device and model pruning method and device

Abstract

The invention provides a pruning threshold determination method and device and a model pruning method and device. The pruning threshold determination method includes: aiming at a current convolution layer of a preset model, determining at least one convolution kernel index position corresponding to convolution kernel combination of the current convolution layer, wherein the current convolution layer refers to any convolution layer of the preset model; obtaining a cumulative distribution function according to a weight value at each convolution kernel index position; taking a preset model compression ratio as a dependent variable, substituting into the cumulative distribution function, and determining an obtained numerical value as a pruning threshold of the current convolution layer. Sincethresholds applicable to any layers of the model are determined on the basis of weight distribution conditions of the corresponding layers of the model, pruning effect optimization is benefited.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a method and device for determining a pruning threshold, and a model pruning method and device. Background technique [0002] With the development and maturity of artificial intelligence technology, more and more deep neural network models are designed, trained, and deployed to different application scenarios. In order to reduce the computational complexity of the model, the trained neural network can be compressed by pruning. Among them, the threshold method is the most commonly used pruning strategy to cut off unimportant connections between neurons in the model, so as to reduce the complexity of the model. [0003] Currently, a threshold can be predetermined and fixed. A fixed threshold can be applied to all layers of the model. Usually, when the threshold is too large, the compression factor increases but the loss of accuracy also increases; if the threshold is too small...

AI Technical Summary

Problems solved by technology

Usually, when the threshold is too large, the compression factor increases but the loss of accuracy also increases; if the threshold is too small, the compression effect is not obvious

[0004] Because the weight distribution of different layers of the model is different, the existing implementation of fixed threshold is not conducive to ensuring the pruning effect

Images