Methods and Devices for Performing Operations in the Convolutional Layer of Convolutional Neural Networks

摘要

PROBLEM TO BE SOLVED: To provide a method for performing a calculation in a folding layer of a folding neutral network which improves the use efficiency of a channel, can reduce an occupation amount of a cache memory, and can improve calculation efficiency, and a device.;SOLUTION: A method includes: a step for padding unfolded feature data which are provided to a folding layer while following a padding method which is designated by the folding layer; a step for creating folded feature data by folding the padded and unfolded feature data at least at one dimension of a width and a height; a step for creating one or plurality of folded folding kernels corresponding to an original folding kernel by folding the original folding kernel at the folding layer at least at one dimension; and a step for performing a folding calculation to the folded feature data by using one or a plurality of the folded folding kernels.;SELECTED DRAWING: Figure 1;COPYRIGHT: (C)2019,JPO&INPIT