A multi-core implementation method of inception structure

Abstract

A multi-core implementation method of the Inception structure, the steps are: S1: According to the parameters such as the number of cores M of the multi-core vector processor and the number N of all convolution cores in a single layer in the Inception structure, to evenly allocate M cores. Each core can be The amount of calculation completed at the same time; S2: N / M convolution kernels are sequentially transferred from DDR to the in-core vector memory bank AM of M kernels, and the output results of the previous level are placed in DDR; S3: M kernels At the same time, the calculation of N convolution kernels is completed, and N output feature maps are obtained at the same time; S4: The output results are transferred from the M kernels to the DDR through DMA, and the M kernels simultaneously load the convolution kernel data of the next level; S5: Broadcast data from DDR to M cores, and complete the calculation of all output feature maps of this stage; S6: Repeat the steps until the calculation of all stages in the Inception structure is completed. The invention has the advantages of simple implementation, convenient operation, can improve the parallelism of the multi-core processor, accelerate the calculation speed of the GoogleNet network model, and improve the calculation efficiency of the multi-core processor.

Description

technical field [0001] The present invention mainly relates to the fields of artificial intelligence, machine learning, and convolutional neural network, and in particular refers to a multi-core implementation method of an Inception structure, which is a multi-core implementation method applied to accelerating GoogLeNet networks. Background technique [0002] The AlexNet model is an important network model in the field of convolutional neural networks. It won the first place in the ILSVRC competition in 2012. Since the AlexNet model regression has added a fully connected layer, the number of parameters reaches 60M. If the hidden layer in the network is increased and the number of layers of the neural network is deepened to further improve the recognition success rate of the network, it will not only increase the total number of parameters in the network, but also increase the number of parameters in the network. Increasing the amount of calculation is also extremely prone to...

AI Technical Summary

Problems solved by technology

Since the AlexNet model regression has added a fully connected layer, the number of parameters reaches 60M. If the hidden layer in the network is increased and the number of layers of the neural network is deepened to further improve the recognition success rate of the network, it will not only increase the total number of parameters in the network, but also increase the number of parameters in the network. Increasing the amount of calculation is also extremely prone to overfitting

Images