An acceleration method for realizing sparse convolutional neural network inference for hardware

Abstract

The invention discloses an acceleration method for realizing sparse convolutional neural network inference for hardware. The method comprises a grouping pruning parameter determination method facing asparse hardware acceleration architecture, a grouping pruning training method for sparse hardware acceleration architecture and a deployment method for forward inference of a sparse convolutional neural network. determining the packet length and the pruning rate of packet pruning according to the number of multipliers in the hardware architecture; based on the magnitude cutting mode, cutting weights except the compression rate; The network accuracy and compression rate after pruning are improved through an incremental training mode, the weight and index parameters of a non-pruning position are saved after the pruned network is finely adjusted, the network is sent to a computing unit under a hardware architecture, and the computing unit obtains the activation value of the packet length atthe same time to complete sparse network forward inference. The pruning parameters and the pruning strategy of the algorithm level are set based on the hardware architecture, the logic complexity of the sparse accelerator is reduced, and the overall efficiency of forward inference of the sparse accelerator is improved.

Description

technical field [0001] The invention belongs to the technical field of electronic information and deep learning, and in particular relates to an acceleration method for realizing sparse convolutional neural network inference for hardware. Background technique [0002] A neural network model is a mathematical expression of a biological neural network learning system. In recent years, with the strengthening of computing power and the generation of large-scale data sets, neural network models have been increasingly used in machine vision fields such as image classification and object detection. [0003] However, from the perspective of using the neural network model to solve problems, people often tend to design a deeper and larger convolutional neural network (CNN for short) to collect more data in order to obtain better results. However, as the complexity of the model increases, the number of model parameters increases, the scale of the model and the floating point numbers r...

AI Technical Summary

Problems solved by technology

Many pruning algorithms have been proposed, but they mainly focus on the number of weights to be pruned, and rarely consider the complexity of deploying the pruned overall network on the ASIC or FPGA accelerator architecture

When the pruned network runs on a hardware accelerator platform, problems such as internal buffer misalignment and load imbalance will occur, making the entire accelerator architecture inefficient

Images