Dependencies

Honestly, I wasn't using an environment for development because all dependencies were already installed on my system and this was faster. In a production situation I'd use pipenv, but right now let me just list dependencies by saying I am using Python 3.9 and rely on latest versions of:

• matplotlib (plotting)
• numpy (plotting)
• PyTorch (optimization)
• nltk (grammar generation)
• scikit-learn (train and test splits)
• gin (config management)
• click (config management)

Realistically only a few of these libraries are strictly necessary.

Running the code

Code relies on Gin for config management. See config/ for example configurations.

Generating the dataset

Use generate.py with a corresponding grammar. See grammar/ for examples:

python generate.py --config config/gen_config.gin

Running training / testing

Call run_traintest.py with the corresponding parameters:

python run_traintest.py --config config/traintest_config.gin train
python run_traintest.py --config config/traintest_config.gin test

Plotting

If you run the training loop with multiple seeds, use plotting.py to generate train/dev loss traces on the output folder like so:

python plotting.py --out_dir out/