/rnnlm_chainer

A Fast RNN Language Model by Chainer

Primary LanguagePythonMIT LicenseMIT

RNN Language Model by Chainer

This is a fast implementation of an RNN language model (RNNLM) by Chainer. This repository is derived from the Chainer example for RNNLM in PTB.

The network architecture is almost same as the "Medium" model in the paper, "Recurrent Neural Network Regularization" by Wojciech Zaremba, Ilya Sutskever and Oriol Vinyals. You can train an RNNLM in 1 miniute per epoch, with backprop length of 35 and batchsize of 20.

How to Run

python -u train.py -g 0 --dataset ptb

Datasets

  • PennTreeBank (ptb)
  • Wikitext-2 (wikitext-2)
  • Wikitext-103 (wikitext-103)

For wikitext, run prepare_wikitext.sh for downloading the datasets.