Skip to content

Latest commit

 

History

History
15 lines (11 loc) · 687 Bytes

README.md

File metadata and controls

15 lines (11 loc) · 687 Bytes
Raw

PytorchRnnLM

Language model using an RNN in PyTorch. Uses the Wiki-Text-2 long term dependency dataset: https://www.salesforce.com/products/einstein/ai-research/the-wikitext-dependency-language-modeling-dataset/

Demonstrates

  1. How to use tied-embedding weights as described by: https://arxiv.org/abs/1608.05859
  2. How to hack PyTorch to use PackedSequence and word embeddings without intermediary padding.

Usage

Download the dataset and unzip file contents into the "wikitext-2" folder. Run "main.py" with "--help" to see possible command line arguments. Default arguments train the model on a mid-class GPU in 10 minutes to get the perplexity score of 135 on the test-set.