NHERI Computational Symposium AI Training

Hands-on training for the NHERI Computational Symposium — running modern AI workflows on DesignSafe with the DesignSafe API (dapi).

CLIPSeg-debris on DesignSafe

CLIPSeg-debris is a text-prompted, 3-class segmentation model that maps hurricane debris (no / low-density / high-density) in post-event aerial imagery. The chapter notebook runs it end-to-end on DesignSafe:

1. Official inference on the published debris dataset (PRJ-6029).
2. A regional debris map for Hurricane Ian (2022) on Estero Island, FL, from NOAA Emergency Response Imagery.
3. Scale-out inference as a GPU HPC job via dapi.
4. Fine-tuning on TACC Vista (GH200), tracked on Weights & Biases or DAPI.

Quick start on DesignSafe (one click)

Clicking the button (uses nbgitpuller) will:

1. Open your DesignSafe JupyterHub session (sign in if you aren't already).
2. Clone — or fast-forward — this repo into ~/MyData/training-ai/ on your account.
3. Open JupyterLab directly on the chapter notebook, ready to run.

Re-click any time to pull the latest version of the code.

The chapter notebook and its utils/ + designsafe_job/ helpers run top to bottom in a DesignSafe JupyterHub session. Parts 3–4 (HPC inference and Vista fine-tuning) need a TACC allocation; Parts 1–2 run without one. Weights & Biases is optional — training can be tracked through DAPI instead.

Authors

• Krishna Kumar, University of Texas at Austin
• Kooshan Amini, Rice University
• Jamie Ellen Padgett, Rice University