arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 1503.02852cs.NEcs.LG
收藏

GPU上广义LSTM类RNN的单流并行化

Single stream parallelization of generalized LSTM-like RNNs on a GPU

Seoul National University(首尔国立大学)

Kyuyeon Hwang, Wonyong Sung

摘要:

递归神经网络(RNN)在处理序列数据方面表现出出色的性能。然而,他们遭受从长的训练时间,这需要并行实施的训练过程。RNN的训练算法的并行化是非常具有挑战性的,因为内部递归路径在两个不同的时间框架之间形成依赖关系。在本文中,我们首先提出了一种广义的基于图的RNN结构,它涵盖了最流行的长短期记忆(LSTM)网络。然后,我们提出了一种并行化方法,通过分析图结构自动探索任意RNN的并行性。实验结果表明,即使使用单个训练流,所提出的方法也表现出很大的速度提升,并且当与多个并行训练流结合时,进一步加速了训练。

英文摘要:

Recurrent neural networks (RNNs) have shown outstanding performance on processing sequence data. However, they suffer from long training time, which demands parallel implementations of the training procedure. Parallelization of the training algorithms for RNNs are very challenging because internal recurrent paths form dependencies between two different time frames. In this paper, we first propose a generalized graph-based RNN structure that covers the most popular long short-term memory (LSTM) network. Then, we present a parallelization approach that automatically explores parallelisms of arbitrary RNNs by analyzing the graph structure. The experimental results show that the proposed approach shows great speed-up even with a single training stream, and further accelerates the training when combined with multiple parallel training streams.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑