Neural network accelerated training method based on eight-bit integer

The invention provides a neural network accelerated training method based on eight-bit integers, and the method comprises the steps: obtaining an eight-bit integer data flow, and building an eight-bit integer matrix based on the eight-bit integer data flow through a preset neural network; performing quantization block division on the eight-bit integer matrix to generate a plurality of quantization blocks; and block-by-block quantization calculation is carried out on a linear layer and a nonlinear layer of the neural network through comprehensive quantization training based on the quantization block, a calculation result is generated, and neural network acceleration training is completed. The problem that an existing network model is high in training cost and long in period is solved. 本发明提供一种基于八位整数的神经网络加速训练方法，包括：获取八位整数数据流，通过预设的神经网络基于所述八位整数数据流建立八位整数矩阵；对所述八位整数矩阵进行量化块划分，生成多个量化块；基于所述量化块通过全面量化训练对神经网络的线性层和非线性层进行逐块量化计算，生成计算结果，完成神经网络加速训练。本发明解决了现有网络模型训练成本高、周期长的问题。