Message Passing Graph Neural Network with Edge Updates in Jraph.
This code implements a graph neural network with the architecture described in https://arxiv.org/pdf/1806.03146.pdf "Neural Message Passing with Edge Updates for Predicting Properties of Molecules and Materials"
All this code is experimental, run at your own risk!
Only tested on Python 3.7.3
See requirements.txt. Only GPU version of JAX has been tested.
If any other libraries are missing, just pip install them, the above list might not be complete.
We use:
- Optax for the training optimizer.
- Jraph for the graph neural network.
- Haiku for the fully connected neural networks (used to compute edge/message updates and for the readout function).
At the moment this can be run with two different datasets: QM9 and aflow.
The QM9 dataset is pulled from spektral and converted into graphs. To get the QM9 dataset run datahandler_QM9.py. You might have to specify and make output file by hand.
The aflow dataset is just a small testset of materials pulled directly with the alfow API with a json response. To do this first run datapuller.py, you may have to make a directory "aflow". Then run datahandler.py to convert the raw data from the pull into graphs.
To train a model run train.py, this defaults to the QM9 dataset, but does not automatically pull it. You have to specify config directory, where parameters are pulled and working directory, where results are stored.
The config directory can be one of the two files in configs.
Only validated on NVIDIA Quadro RTX 4000 with 8GB of VRAM. Quadro P400 with 2GB of VRAM runs out of memory.