Neural networks for protein spherical representations, based on the Deepfold implementation by Boomsma and Frellsen. (W. Boomsma, J. Frellsen, Spherical Convolutions and their application in molecular modelling, Conference on Neural Information Processing Systems (NIPS) 30, 2017.)
This is my project of a Convolutional Neural Network (CNN) on a spehrical coordinate system. It tries to stay as close to state-of-the-art CNN implementation best-practices as possible, as well as making use of all the options available on the tensorflow framework for training, testing and inference.
The main differences between this and DeepFold are Tensorboard usage with summaries for training performance assessment, full tensorflow checkpoint loading for training continuation with epoch and step as graph variables and finally a Sparse Tensor implementation of the data loader for faster training. Slight tweaks to the models were added as well, in order to bring these closer to state-of-the-art code, such as batch normalization.
Part of the reasons that lead to this project was to develop my own default project organization, as model development in tensorflow tends to be a bit convoluted. Having a well organized project allows for easy expansion and experimentation.
The current one is heavily influenced by Morgan Giraud's implementations, explained here, as well as other random discussions online about best practices.
Training on the dense matrix form such as the one from Deepfold data leads to slow training times. This poor performance is mostly due to slow transfer times between the CPU and GPU when the matrix is passed in the dense form, leading to a lot of GPU waiting time between steps. I've implemented a sparse matrix representation feeding mechanism through the feed dict feature of tensorflow, leading to almost zero GPU lag time and constant 100% usage on a single Titan X (12GB RAM) GPU.
This is a work-in-progress project, geared towards my own experimentation with tensorflow and its capabilities. As such, some improvements are in the works whenever I have the time to explore and include them, such as queues for asynchronous data access and multi-gpu training with overall loss calculation. Also, the code might not be the cleanest at all times. :)