SchurNet

Official codebase for SchurNet

SchurNet

Package dependencies

1. Install ptens dev4 branch: https://github.com/risi-kondor/ptens/tree/dev4
2. Install remaining packages listed in requirements.txt

Current config structure:

To create your own experiment, under config/ folder, add an yaml file named experiments.yaml with the following content:

defaults:
  - experiment_settings: <file_name>
  - dataset: <file_name>
  - training_hyperparameters: <file_name>
  - model: <file_name>

We have uploaded the configs for the experiments in the paper, which includes the ablation studies on TU Dataset, experiments on ZINC, ogbg-molhiv, ogbg-moltox.

Training models in single GPU

To train a model in a specific config file, please refer to the following command:

    python main_batch.py --config_name=<config_file_name>

To debug hydra (in case the configs are incorrectly configured):

    HYDRA_FULL_ERROR=1 python main_batch.py --config_name=<config_file_name>

To add a new model, create the model under model/ folder. Then include in the config/model folder both the model name, file path, and input parameters needed to initialize the model.

Training models in multiple GPUs

To train a model in a specfic config file using multiple gpu, please refer to the following command:

./launch_multi_gpu.sh <config_file_name>

TODO:

1. compatible with dev5 branch of ptens package
2. checkpoints for different experiments