Setup

Add src directory to PYTHONPATH and install dependencies.

export PYTHONPATH='.'
pip install -r requirements.txt --use-feature=2020-resolver

Experiments

After installing your dependencies, you can run the following scripts to reproduce our experiments:

Preprocessing

Pulls COGS dataset, and preprocess sequential outputs into graph outputs under the cogs/data folder:

bash setup_cogs.sh

You can change the TARGET_DIR var in the above script to save the data in a different folder. Alternatively, for CFQ, you can run bash setup_cfq.sh which will perform the same and save data under cfq/data.

COGS: Transformer Baseline (hyperparameters are set as default):

python src/train.py

COGS: best strongly-supervised Transformer LAGr:

python src/train.py --from_config cogs/strongly_sup_hyperparams.yaml

COGS/CFQ: Best weakly-supervised Transformer LAGr:

For COGS:

python src/train.py --from_config cogs/weakly_sup_hyperparams.yaml

For CFQ, run the following to train a model on the MCD1 split.

python src/cfq_train.py --from_config cfq/weakly_sup_hyperparams.yaml --split mcd1