Capsar/dat

Combination of Distributed Adversarial Training and JointSpar-Lars to experiment the effects of sparsifying gradients and their computation on Distributed Adversarial Training.

DAT with JointSpar

This repo is a combination Distributed Adversarial Training to Robustify Deep Neural Networks at Scale and Communication-efficient Distributed Learning for Large Batch Optimization.

The repository is adjusted such that it can be run on Google Cloud with Multiple Node and Multiple GPUs, the details are in the GCloud Interface Jupyter Notebook

Plots of the results can be found in the Plot Experiments Jupyter Notebook