CNN quantification method based on low-precision floating-point number, forward calculation method and device

Abstract

The invention discloses a CNN quantification method based on a low-precision floating-point number and a forward calculation method and device, and relates to the field of deep convolutional neural network quantification. The quantification method comprises the following steps that a low-precision floating-point number representation form MaEb of a network is defined; in the process of optimizingthe low-precision floating-point number representation form, the optimal low-precision floating-point number representation form and the optimal scaling factor corresponding to the minimum mean squareerror value are obtained by changing the scaling factor, changing the combination of a and b and calculating the weight and activation value mean square error before and after quantification; based on the low-precision floating point number representation form and the optimal scale factor, single-precision floating points are quantified into low-precision floating points. According to the method,the low-precision floating-point number representation form MaEb is calculated and used, so that the accuracy of network quantization is ensured under the condition that retraining is not needed; under the condition that the accuracy is guaranteed, the acceleration performance of a customized circuit or a non-customized circuit is greatly improved, the customized circuit is an ASIC or an SOC, andthe non-customized circuit comprises an FPGA.

Description

technical field [0001] The invention relates to the field of deep convolutional neural network quantization, in particular to a CNN quantization method, a forward calculation method and a device based on low-precision floating-point numbers. Background technique [0002] In recent years, the application of AI (Artificial Intelligence, artificial intelligence) has penetrated into many aspects, such as face recognition, game battle, image processing, simulation, etc., although the processing accuracy has been improved, but because the neural network contains many layers and A large number of parameters require a very large calculation cost and storage space. In this regard, technicians have proposed a neural network compression processing scheme, that is, by changing the network structure or using quantization and approximation methods to reduce network parameters or storage space, and reduce network cost and storage space without greatly affecting neural network performance. ...

AI Technical Summary

Problems solved by technology

[0004] Although the current technology improves the quantization and improves the quantization accuracy, there are still several limitations: 1) For quantized deep convolutional neural networks (the number of convolutional layers / fully connected layers exceeds 100 layers), retraining is required to Guaranteed accuracy; 2) Quantization needs to use 16-bit floating-point numbers or 8-bit specific points to ensure accuracy; 3) Under the premise of not using retraining and ensuring accuracy, the current technology can only achieve two at most in one DSP. multiplication operations, resulting in lower acceleration performance on the FPGA

Images