Anna-Bielawska/automating-science

The Automating Science course project repository regarding GNN molecular property prediction.

Molecule generation loop benchmark

A benchmark for testing of different molecule generation loops.

Loop implementations are tested on the ZINC dataset.

Usage

To start the benchmark, you need to provide the necessary configuration parameters through the command line.

The benchmark is started by running the main.py script with the desired configuration parameters.

Currently, the required configuration parameters are:

• model - the model to use for the benchmark
• training.loop - the loop to use for the benchmark

Example command to run the benchmark:

python main.py model=GraphNeuralNetwork training.loop=gnn_loop pretraining.enabled=True _repeat=1

Configuration

Possible configurations and command line arguments are located in the config directory.

The configuration is implemented using the Hydra library through Python dataclasses.

Custom model and loop implementations

To add a custom model or loop implementation.

• Create a new file in the models, loops or embeddings directory with the implementatin of the model or loop.
• Create a new configuration file in the config directory with the new model or loop, following the existing configuration files as an example.
• Register the model, loop or embedding with a @register_model, @register_loop or @register_embedding decorator.

Results analysis

The benchmark results are saved in the outputs directory. For analysis of the results, you can use the results_analysis.ipynb Jupyter notebook.