fred-xue / annotated-transformer

This fork of the Annotated Transformer works with PyTorch 1.3

Code for The Annotated Transformer:

http://nlp.seas.harvard.edu/2018/04/03/attention.html

Authors: Alexander Rush (@harvardnlp or [email protected]), with help from Vincent Nguyen and Guillaume Klein.

Modified by Michel Galley to make it work with PyTorch 1.3.

Setup

Assuming you have PyTorch 1.3 already installed, please run this before loading the notebook:

pip install -r requirements.txt
wget https://s3.amazonaws.com/opennmt-models/iwslt.pt
python -m spacy download en
python -m spacy download de

Then, Run All cells should work in one pass without any error.

Changes

The code differs from the original Annotated Transformer in the following ways:

• Runs on PyTorch 1.3 (original code was for version 0.3);
• As the original en-de OpenNMT doesn't load in recent versions of PyTorch, the last part of the notebook (attention visualization, etc.) uses an IWSLT rather than WMT model;
• Removed some deprecations when possible;
• Fixed seed for reproducibility;
• Decoding with slightly more difficult examples.

Detailed changes are listed in annotated_transformer.nbdiff.

Credit

If you use this for anything, please give credit to Alexander Rush. Here is a paper you can cite:

@inproceedings{opennmt,
  author    = {Guillaume Klein and
               Yoon Kim and
               Yuntian Deng and
               Jean Senellart and
               Alexander M. Rush},
  title     = {OpenNMT: Open-Source Toolkit for Neural Machine Translation},
  booktitle = {Proc. ACL},
  year      = {2017},
  url       = {https://doi.org/10.18653/v1/P17-4012},
  doi       = {10.18653/v1/P17-4012}
}