Implementation of Deep Tensor Neural Network (DTNN) in PyTorch based on the paper Quantum-Chemical Insights from Deep Tensor Neural Networks originally designed for prediction of energy of a molecule based on the QM9 dataset. Slight modifications have been made on this implementation to accomodate for the multiple target variables of the QM8 dataset.
Take note that there are 2 implementations. One which implements the network
using primarily vanilla Pytorch features, for this take a look at
models/vanilla.py
. The second method implements the network using a message
passing neural network, based of off the torch_geometry
library, look at the
file torch_geom.py
.
For an explanation of the inner workings and explanation of implementations do take a look at the slides directory.