Method for processing convolution neural network

Abstract

The invention provides a method for processing convolution neural network. After inputting input data to a floating pre-trained convolution neural network to generate floating feature maps for each layer of the floating pre-trained CNN model, a statistical analysis on the floating feature maps is performed to generate a dynamic quantization range for each layer of the floating pre-trained CNN model. Based on the obtained quantization range for each layer, the proposed quantization methodologies quantize the floating pre-trained CNN model to generate the scalar factor of each layer and the fractional bit-width of a quantized CNN model. It enables the inference engine performs low-precision fixed-point arithmetic operations to generate a fixed-point inferred CNN model.

Description

technical field [0001] The present invention relates to image processing, in particular to an energy-saving convolutional neural network implementation. Background technique [0002] Due to its outstanding success in the ImageNet competition, Convolution Neural Network (CNN) has become the most popular structure in computer vision processing. A typical pre-trained CNN model requires millions of operations, a large amount of memory space, and several watts of power for a single inference operation. Limited computing resources and storage space have become the main obstacles to implementing CNN on Internet of things (IoT) or portable devices. [0003] There are three main challenges in developing new CNN accelerators: [0004] Spatial data transfer using limited storage memory: Due to limited memory (<320KB SRAM) in IoT devices, real-time artificial intelligence (AI) applications cannot accept off-chip memory such as The delay of large data transfers between dynamic rand...

AI Technical Summary

Problems solved by technology

[0007] CNN requires a lot of arithmetic operations, so it cannot be implemented on low-power devices

Images