Large-Scale Transfer Learning for Natural Language Generation
Team
- Sergey Golovanov - sergey.golovanov@neuromation.io
- Rauf Kurbanov - rauf.kurbanov@neuromation.io
- Sergey Nikolenko - snikolenko@neuromation.io
- Kyryl Truskovskyi - kyryl@neuromation.io
- Alexander Tselousov - al.tselousov@gmail.com
- Thomas Wolf - thomas@huggingface.co
How to run
Unzip BPE vocabulary files into ./parametes folder and save checkpoint into
./checkpoints folder or use scripts (see below).
The easiest way to prepare environment is to run script prepare_environment.sh.
After that docker container with retrieval server must be run in demon mode and
image with transformer_chatbot must be built. Run scripts from the root folder of this repository.
After preparations metrics can be evaluated with corresponding docker_*.sh scripts or
*.py scripts can be used during interactive container run.
List of used python modules is in requirements.txt. Also pytorch=1.0 is used.
Run all experiments on the platform
All experiments were run on Neuromation platform
./platform/train.sh