LSTM model optimization method, accelerator, device and medium

Abstract

The invention discloses an LSTM model optimization method, an accelerator, a device and a medium. The method comprises the following steps: acquiring a weight matrix of a pruned LSTM network; based on the weight matrix, obtaining the weight sparseness of the weight matrix; obtaining the sparseness of an input sequence; based on the weight sparseness and the sparseness of the input sequence, judging an operation mode; and if the operation mode is judged to be a sparse operation mode, calculating the input sequence according to the sparse operation mode. The invention aims to improve the energy efficiency ratio of the LSTM hardware accelerator, and is simple to implement and easy to deploy.

Description

technical field [0001] The invention relates to the field of computer hardware acceleration, in particular to an LSTM model optimization method, accelerator, device and medium. Background technique [0002] Recurrent Neural Network (RNN) based on Long Short-Term Memory (LSTM) is a neural network model for processing sequence data. It is widely used in cognitive fields, such as speech recognition, behavior recognition and natural language processing. But in the actual engineering application practice, it faces many problems. Traditional computing platforms cannot carry such a large amount of data calculation of LSTM. In embedded applications, especially in areas with extremely high latency requirements such as autonomous driving, the LSTM model itself has a huge amount of parameters, as well as a large amount of training data and inference test data, resulting in not only computing The complexity is high, and the power consumption of the computing platform is also very lar...

AI Technical Summary

Problems solved by technology

Traditional computing platforms cannot carry such a large amount of data calculation of LSTM

In embedded applications, especially in areas with extremely high latency requirements such as autonomous driving, the LSTM model itself has a huge amount of parameters, as well as a large amount of training data and inference test data, resulting in not only computing The complexity is high, and the power consumption of the computing platform is also very large

It is difficult to deploy models in embedded devices with extremely high power consumption requirements

[0003] For the hardware-accelerated calculation of the sparse LSTM model, the industry introduced the Delta algorithm to construct and mine the sparsity of the sequence data by using the numerical similarity of the sequence data, and reconstruct the LSTM model and accelerate the algorithm. However, this method Limited by the time dependence of sequence data, the input data at adjacent times needs to have a high similarity, and the scope of application has obvious limitations, and the implementation of hardware acceleration for Delta-based LSTM model reconstruction is complicated, which is not conducive to hardware deployment

Images