requirements: atomic-images, tensorflow 2.0
This installation guide expects the user to understand pip and have it installed. This package depends on
atomic-images available here.
Clone atomic-images using git clone git@github.com:UPEIChemistry/atomic-images.git, then checkout the tf_2.0 branch
by using
git checkout tf_2.0. Once this branch is checked out, use pip install -e ./atomic-images to install the package.
To install tensor-field-networks, start by cloning the repo with
git clone git@github.com:UPEIChemistry/tensor-field-networks.git,
followed by using pip: pip install -e ./tensor-field-networks. The setup.py script contained in this package
should install tensorflow 2, numpy, and any other 'official' dependencies. Be sure to install tensorflow-gpu==2.0.0
and CUDA/cudNN if you intend to use this code on a GPU (which is recommended for the performance boost).
Tensor Field Networks (TFN) are Rotationally Equivariant Continuous Graph Convolution Neural Networks which are capable of inputing continuous 3D point-clouds (e.g. molecules) and making scalar, vector, and higher order tensor predictions which rotate with the original input point-cloud (Thomas et. al., 2018).
Ignoring the continuous convolution part, this means that TFNs are capable of knowing when an image has been rotated, something vanilla convolution nets are not capable of. For example, a traditional conv. net trained to recognize cats on non-rotated images would not identify a cat in the second picture:
While TFNs will still identify a cat in the rotated image, trained only on images in a single orientation. To see a
demonstration of this equivariance, and a further explanation of TFNs, checkout the Jupyter notebook located in the
tutorials directory. If the user is not familiar with using Jupyter notebooks, they can read up on them
here.