Welcome to the first global, virtual TRANSFORM!The Software Underground is excited to bring you lots of interesting and useful sessions on the digital subsurface. The schedule is still taking shape, but this event is a bit different from most conferences: the sessions are all fully participatory and interactive. Get ready for a conversation!
If you are viewing the schedule on the web, you can select the timezone you wish to see in the menu on the right. In the mobile app, times are shown in UTC.
As a starting point I assume you are familiar with basic sklearn model pipelines and the “shape” of NN training code. We’ll be using pytorch for that bit.
UPDATED
Then the tutorial will focus on Ray and its libraries. Exploring some of the training code patterns used, and the Ray/Raytune API. Once your code can be run by Ray then it’s easy to use ray features and train models on a multi core machine though to a multinode gpu cluster without much change to that core code. Ray does the heavy lifting.
So we’ll step through various ray features taking a more verbose and relaxed path then some of the ray docs, and (hopefully) with some examples relevant to the subsurface domain.
We’ll spend most of our time looking at SOTA Hyperparameter optimisation with Raytune, Ray itself and [hopefully] plpaying with a Ray Cluster.
But we’ll also at least pay homage to RayRLib (Reinforcement Learning) RaySGD (Distributed Training) and look at the very recent addition Ray Serve.
Some parts of the tutorial you will be able to run on a local machine. More if you have a GPU available. Some parts may involve launching AWS clusters and that part will probably be “follow along” only.