README.md

Encoder-based models

Setup

All encoder-based models use the same environment. Instructions for creating this enivornment are as follows. You can see our exact creation scripts at scripts/setup.

Create the conda env (named coref_env):

conda create -y -n coref_env python=3.9 cudatoolkit=11.3.1 --override-channels -c conda-forge -c nvidia
conda activate coref_env
conda install -y pytorch=1.12.1 torchvision==0.13.1 torchaudio==0.12.1 --override-channels -c nvidia -c pytorch -c conda-forge -c defaults

conda install -y scikit-learn pandas matplotlib hydra-core --override-channels -c conda-forge
conda install -y transformers==4.27.2 datasets==2.10.1 sentencepiece pytorch-lightning==2.0.0 --override-channels -c conda-forge
conda install -y jsonlines wandb --override-channels -c conda-forge

pip install udapi
pip install nltk

Activate the environment and login to huggingface and wandb:

conda activate coref_env

wandb login
huggingface-cli login

Training

To train a respective model, use experiments/train_*/train.py.

Inference

Use inference/inference_*.py.