rnn : Recurrent Library for Torch7
Nicholas Léonard Element Inc. New York, NY \AndSagar Waghmare Element Inc. New York, NY \AndYang Wang Element Inc. New York, NY \AndJin-Hwa Kim Seoul National University Seoul, Republic of Korea
Abstract
The rnn package provides components for implementing a wide range of Recurrent Neural Networks. It is built withing the framework of the Torch distribution for use with the nn package. The components have evolved from 3 iterations, each adding to the flexibility and capability of the package. All component modules inherit either the AbstractRecurrent or AbstractSequencer classes. Strong unit testing, continued backwards compatibility and access to supporting material are the principles followed during its development. The package is compared against existing implementations of two published papers.
中文速览
循环神经网络（Recurrent Neural Network, RNN）的实现历来繁琐，研究者往往需要手动克隆模块、共享参数，并自己编写序列遍历代码。作者在 Torch 框架下开发了 rnn 工具包，通过三轮迭代逐步引入 Recurrent、LSTM、Sequencer、Repeater、RecurrentAttention、Recursor 等可组合模块，将递归结构封装成与 nn 包完全兼容的标准接口，让用户像搭积木一样构建各类 RNN。该包在两篇已发表论文的复现实验中均取得了与原文一致的结果，验证了其正确性与通用性。它的意义在于大幅降低了实现复杂 RNN 架构的门槛，同时保持了良好的向后兼容性和可扩展性，为深度学习社区提供了一个经过充分测试、文档完善的循环网络开发基础设施。
原文 arXiv:1511.07889;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1511.07889v2