entry_points.md

All of these are executed from the src/ directory.

1. python pipeline_pre.py 1, which

• Reads training data from RAW_DATA_DIR
• Adds enhanced information to new csv files in PROCESSED_DATA_DIR
• This can take about an hour to run, mostly due to external libraries (rdkit and openbabel).

2. python pipeline_pre.py 2, which

• Reads enhanced information from new csv files in PROCESSED_DATA_DIR
• Creates loaders, which are pickled into PROCESSED_DATA_DIR
• This can take 60-90 minutes to run, mostly creating the data loaders

3. python train.py --cuda [--multi-gpu] [--batch_size 60] [--d_model 700] [...], which

• Trains a graph transformer model with specified parameters and for designated number of epochs
• More options can be found via the -h flag. Some commonly used flags are --d_model (the embedding dimension), --dropout (fully connected layer dropout rate), and --n_layer (number of transformer layers). You can also specify on whether to use 80% or 100% of the training data with --mode (if 80%, the rest of the data will be used as a validation set).

4. python predictor.py, which would

• Read list of models from CONFIG_DIR/models.json
• Read those models from MODEL_DIR
• Use them to make predictions, using test.csv in RAW_DATA_DIR (can take ~15 min each for the Kaggle test set)
• Write the predictions of each model to SUBMISSION_DIR
• Ensemble those into a new prediction csv in SUBMISSION_DIR

5. python predictor.py fast, which would only

• Read the predictions of each model from SUBMISSION_DIR
• Ensemble those into a new prediction csv in SUBMISSION_DIR