training

Training

Example training to test the performance of model implementations of kgcnn per datasets. The training scripts are called via:

python3 train_node.py --hyper hyper/hyper_cora.py --category GCN
python3 train_graph.py --hyper hyper/hyper_esol.py --category GIN

Where hyper_esol.py stores hyperparameter and must be in the same folder or a path to a .py. Alternatively, also a .yaml, .yaml or .json file can be loaded in place of the .py file. The python file must define a hyper attribute as described below. However, note that not all models can be trained on all datasets and that not all models have proper default hyperparameter here. You can check previous runs in the result folders named after each dataset and their hyperparameter and output files.

There is a make_config_training.ipynb jupyter notebook to demonstrate how the .py config file can be set up and provide further information on how it is structured. In short the config file contains a python dictionary of the form:

hyper = {
    "info":{ 
        # General information for training run
        "kgcnn_version": "4.0.0", # Version 
        "postfix": "" # Postfix for output folder.
    },
    "model": { 
        # Model specific parameter, see kgcnn.literature.
    },
    "data": { 
        # Data specific parameters.
    },
    "dataset": { 
        # Dataset specific parameters.
    },
    "training": {
        "fit": { 
            # serialized keras fit arguments.
        },
        "compile": { 
            # serialized keras compile arguments.
        },
        "cross_validation": {
            # serialized parameters for cross-validation.  
        },
        "scaler": {
            # serialized parameters for scaler.
            # Only add when training for regression.
        }
    }
}

Furthermore, you could also have a dictionary of models as hyper={"GCN": {...}, ...} which each model has a config dictionary as above. The category command line argument is used to select which category aka model/dataset settings to choose.

If a python file is used, also non-serialized hyperparameter for fit and compile can be provided. However, note that the python file will be executed, and a serialization after model fit may fail depending on the arguments.