Adaptive distributed parallel training method for neural network based on reinforcement learning

Abstract

The invention discloses an adaptive distributed parallel training method for a neural network based on reinforcement learning, and provides an optimal solution for segmentation and scheduling of a large-scale complex neural network. Firstly, the influence of a neural network model structure and calculation attributes on execution performance is analyzed, on this basis, performance factors including calculation cost, communication cost, memory utilization rate and the like are extracted, a multi-dimensional performance evaluation model capable of comprehensively reflecting distributed training performance is constructed, and comprehensive performance of a parallel strategy is improved; secondly, self-adaptive grouping of operators is realized according to attribute characteristics of the operators by utilizing a feed-forward network, the degree of parallelism is determined, and end-to-end strategy search is realized while the search space is reduced; and finally, based on importance sampling, a near-end strategy gradient iteration optimization reinforcement learning model is adopted, an optimal segmentation and scheduling strategy is searched, the strategy network offline learning capability is expanded, and algorithm stability, convergence rate and strategy search performance are improved.

Description

technical field

[0001] The invention relates to a neural network adaptive distributed parallel training method based on reinforcement learning, which provides an optimal model parallel training scheme for large-scale complex neural networks. Background technique

[0002] In recent years, benefiting from the development of AI algorithms, hardware computing power, and data sets, deep neural network technology has been widely used in natural language processing, computer vision, and search recommendation. As these fields continue to iteratively develop larger-scale and more complex neural networks, it is difficult for "Moore's Law" to match the computing needs, and a single device can no longer support large-scale deep network training. Therefore, it has become a common method to solve large-scale neural network training by researching and dividing the neural network calculation graph, and scheduling the divided network to clusters containing multiple CPUs and GPUs to achieve m...

Images