Reference implementation of the DimeNet model proposed in the paper:
Directional Message Passing for Molecular Graphs
by Johannes Klicpera, Janek Groß, Stephan Günnemann
Published at ICLR 2020.
This repository contains a notebook for training the model (train.ipynb) and for generating predictions on the test set with a trained model (predict.ipynb). It also contains a script for training the model on a cluster with Sacred and SEML (train_seml.py).
For faster experimentation we offer a set of pretrained models, which you can find in the pretrained folder. On average, these models outperform the results reported in the paper by 3% (see table below).
This difference is due to slightly different training and initialization in TF2 and to using orthogonal Glorot initialization in the output layer for the targets alpha, R2, U0, U, H, G, and Cv, while using zero initialization for Mu, HOMO, LUMO, and ZPVE. The paper used the exact same architecture and hyperparameters in all experiments. It only used zero initialization for the output layer.
Training the original DimeNet architecture is rather slow (around 20 days for 3M steps on an Nvidia GTX 1080Ti). We are currently looking into ways of improving this. One simple and amazingly effective tweak is exchanging the bilinear layer in the interaction block with a simple Hadamard product, while transforming the 2D basis embedding a_SBF with an MLP instead of a linear layer. This yields roughly a 4x improvement without sacrificing any performance.
The repository uses these packages:
numpy
scipy>=1.3
sympy>=1.5
tensorflow>=2.1
tensorflow_addons
tqdm
Please contact klicpera@in.tum.de if you have any questions.
Please cite our paper if you use the model or this code in your own work:
@inproceedings{klicpera_dimenet_2020,
title = {Directional Message Passing for Molecular Graphs},
author = {Klicpera, Johannes and Gro{\ss}, Janek and G{\"u}nnemann, Stephan},
booktitle={International Conference on Learning Representations (ICLR)},
year = {2020}
}