Offline Translation

Author

• Alexander Wolf

Last Commit Date

• 3/05/18

Notebook Overview

This documented notebook shows how you can easily train a neural translation model using the Tensor2Tensor API.

Supplied Translation Datasets by Tensor2Tensor

All of these datasets are for translating to English or from English by T2T

1. French
2. German
3. Chinese
4. Chzech
5. Macedonian

Adding Your Own Datasets

It is possible to add your own translation dataset to Tensor2Tensor and use their models/ easy deep learning framework if they don't supply the language you are translating to. You can view this link on the Tensor2Tensor API to see how to.

I recommend using Sublime SFTP extension for adding extensions to the library if your running it from a remote server.

Tensorboard

T2T is Tensorboard compatible, and logs tons of important metrics for you automatically, including train/eval loss, BLEU Score, and much more.

• Notes about Tensorboard
  • You need terminal access
  • If running training on a secure server, and you want to view the results on a non-confiured IP address to the server, you can't use local host to ssh in
    • Example proper remote shell for Tensorboard: ssh -L 16006:new_ip:6006 user@server_url.com
    • With this you can access Tensorboard at 127.0.0.1:16006 instead of the default localhost:16006
• To use Tensorboard
  1. Activate you python enviroment in the console which contains tensorflow
     • source PATH_TO_PYENV/bin/activate
  2. Use Tensorboard on your train_dir in T2T, where your model checkpoints are saved
     • tensorboard --logdir PATH_TO_MODEL_train_dir
  3. Go to http://host_ip:6006/ to see your model's training and eval stats

Measuring Translation Accuracy

• Benchmarking good translation models/accuracy is difficult due to the numerous of ways things can be translated
  • The BLEU Score is the most commonly used and practical one
• Similarly different languages can translate more easily to one another, so you should do your research for what a good BLEU score for the language translation pair is.
• During evaluation T2T uses the approximate BLEU Score, not the real used in research papers since it is computationally expensive to compute
  • T2T supplies a shell script for getting the real BLEU score
  • EX: t2t-bleu --translation=translation.en --reference=ref-translation.de

My Results

• I was able to Reach a BLEU score of 40 for English-> French and 33 for French-> English which is considered a state of the art performance
• Used the Transformer model from this research paper by google and model
  • The Transformer architecture is precoded into the library
• Here is a photo of Tensorboard showing the training o an EN -> Fr model that reached a state of the art BLEU score of 40
  • 
• Training/Validation Loss Photo
  • 

To Do Still

• Show how to decode a translation model for big dataset, and few txt examples
• Make notebook for getting real BLEU score on model