BCAI_kaggle_CHAMPS | Bosch solution to CHAMPS Kaggle competition

We use only the train.csv, test.csv, and structures.csv files of the competition. They should be (unzipped and) placed in the data/ directory. All of the commands below are executed from the src/ directory.
The config/models.json file contains the following important keys:.
Very fast prediction: predictor.py fast to use the precomputed results for ensembling.
Ordinary prediction: predictor.py to use the precomputed checkpoints for predicting and ensembling.
Re-train models: train.py to train a new model from scratch. See train.py -h for allowed arguments, and config files for each model for the arguments used.
names: List of the names we will ensemble
output file: The name of the ensembled output file
num atom types, bond types, triplet types, quad types: These are arguments to pass to the GraphTransformer instantiator. Note that in the default setting, quadruplet information is not used by GTs.
model_dir: The directory in models/ associated with each model. Each directory must have graph_transformer.py with a GraphTransformer class (and any modules it needs); config file with the kwargs to instantiate the GraphTransformer class; [MODEL_NAME].ckpt that can be loaded via load_state_dict(torch.load('[MODEL_NAME].ckpt').state_dict()) (to avoid PyTorch version conflict).