THRIVE-TN demo — command cheat sheet

NSF NAIRR · THRIVE-TN · VALIANT AI Summer School 2026 · Friday Aug 14 · Stevenson Center 4

Your ticket says a model and a number — e.g. voxelmorph 3. Type both, exactly as printed. Anything in <angle brackets> is yours to replace; everything else is typed literally.

Every command below has a copy button. Copy it, paste it, press enter.

---

Demo 1 — log in and claim a graphics card

D1C1

ssh -o ServerAliveInterval=120 <username>@login.delta.ncsa.illinois.edu

Your NCSA username, not your ACCESS ID. Approve the Duo prompt on your phone.

D1C2

mkdir -p ~/lumir && cp -r /projects/bfep/thrive_ai_valiant_2026/LUMIR_demo/scripts/. ~/lumir/

Copies the scripts into your own home folder. One long line — no line break. Safe to run twice: if you already have a ~/lumir from an earlier session, this replaces the scripts in it rather than burying a second copy inside.

D1C3

cd ~/lumir

Everything for the rest of the day is run from here.

D1C4

./gpu.sh

Claims a graphics card. The prompt changes and a banner names the machine, the card and when the session ends — from here on you are running in Illinois.

---

Demo 2 — split the work

100 pairs of brains, shared across the room.

D2C1

./task21.sh <model> <ticket>

Task 2.1, same scan type. Two to three minutes, then a box with three numbers. Nobody types a number anywhere — the board at the front fills itself.

D2C2

./task22.sh <model> <ticket>

Task 2.2, different scan type — same pairs, same anatomy, same right answer. Two to three minutes.

---

Demo 3 — split the question

Everyone trains, everyone at a different setting.

If your session expired over lunch, log in again with D1C1 and claim a card again with D1C4. Then:

D3C1

./task31.sh <model> <ticket> <strength>

Task 3.1, train your own model. <strength> is the number you picked between 0 and 1 — write it on your ticket, you need it again. The script reads it back before it starts, so a slip costs a second rather than ten minutes. Eight to ten minutes, and the == epoch lines should climb.

D3C2

./task32.sh <model> <ticket>

Task 3.2, score your own model on the same pairs. Two things only — your strength is picked up from D3C1, so do not retype it.

---

The two scores

	Means	Which way is good
DSC (Overlap)	how well the anatomy lines up afterwards	higher is better
Folding (NDV)	how much of the bending was physically impossible	lower is better

If something is wrong

• A command fails — put your hand up and keep it up. Don't retype it.
• Your terminal closed, or the session ended — run ./gpu.sh again. It reconnects to the session you already have rather than queueing a second one.
• A run is taking more than twice as long as the times above — tell an instructor.

The four models are voxelmorph, transmorph, vfa and sitreg. Which one you drew is on your ticket, and it is how the four groups find each other.