In this tutorial and the previous one, we have introduced a lot of content on how to write training scripts that run on multiple GPUs with the MNIST dataset. For the rest of this notebook we will return to a more interesting application, the DesignSafe Classifier we created in previous tutorials. As a reminder, this is a dataset from Hurricane Harvey, a category 4 hurricane that hit Texas in August of 2017 and resulted in catastrophic flooding to the Houston metropolitan area. The data set is specifically focused on image classification of homes according to the amount of damage the home received. All images of homes are labeled as C0, C2, or C4 respectively for low, medium or high damage.

For the remainder of this notebook we will do a code walk through for the DesignSafe Classifier that is set up to run on multiple GPUs with torchrun. Additionally this example will incorporate code for model evaluation. The script we create in this notebook will be used in the next tutorial where we cover multi-node training. Let’s dive in!