Acceleration method of convolutional neural network parallelization training

The invention provides an acceleration method of convolutional neural network parallelized training and a mixed-batch idea. The method is applied to a complete machine system composed of a CPU and anFPGA, and mainly solves the problem that under a large-scale convolutional neural network structure, when the FPGA is used for parallelization training of one batch sample, storage space is insufficient, and the method can be applied to image recognition and target detection in the field of computer vision. The above method includes the following steps that 1, in the data preprocessing stage, thesamples of a original training library are randomly rearranged; 2, in the feedforward calculation stage, data is written in shared memory in the form of the batch, based on the parallel processing ofeach layer of the convolutional neural network achieved through an OpenCL language, data of one sample in the batch of the previous layer is randomly read in a first full-connection layer in whole internet, and the output of th