Depth-first data scheduling method, system and equipment based on block convolution

Abstract

The invention belongs to the field of convolutional neural networks, particularly relates to a depth-first data scheduling method, system and equipment based on block convolution, and aims to solve the problems that an existing convolution model calculation method needs to perform calculation layer by layer, and storage of an intermediate result feature map needs to occupy a large amount of memory and is not suitable for being deployed in full-hardware equipment. The method comprises the following steps: dividing an input feature image into a plurality of blocks, calling the blocks one by one to carry out convolution or maximum pooling to generate a next-layer feature map, if the next-layer feature map reaches a preset block size, continuing to call to the next layer to obtain a deeper feature map, and if the next-layer feature map is smaller than the preset block size, returning to the 0-th layer to call until the reasoning process is completed. According to the method, memory consumption caused by storage of a large number of convolution layer intermediate results is avoided, and the reasoning efficiency of the convolution model on full hardware equipment is improved.

Description

technical field [0001] The invention belongs to the field of convolutional neural networks, and in particular relates to a block convolution-based depth-first data scheduling method, system and equipment. Background technique [0002] With the continuous development of deep learning technology, a series of models represented by convolutional neural networks have achieved good results in image classification, object detection and other fields, and have been widely used in daily life. However, the feature maps of each convolutional layer in the convolutional neural network are usually large, and the method of layer-by-layer convolution will occupy a large amount of memory, and the full hardware device usually has limited memory, which makes it difficult for the convolution model to be implemented on the full hardware device. Deployment, to a certain extent, limits the application of convolutional neural networks. In addition, the layer-by-layer convolution method can only per...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the above-mentioned problems in the prior art, that is, the data scheduling of the existing full-hardware equipment needs to convolve the entire image at the same time, and the processing of the image occupies too much memory and is not suitable for deployment in the full-hardware equipment. The present invention provides A depth-first data scheduling method based on block convolution is proposed, including:

Images