Skip to content

Reproduce

This is a starter reproducibility guide. I’ll make the commands and dataset IDs fully exact once the thesis pipeline is locked.

Goal

You should be able to (1) set up the environment reliably, (2) run a precomputed evaluation (coming soon), and (3) later reproduce training + evaluation from tracked configs.

1) Clone the code repo

git clone git@github.com:OAE-PINN-Emulator/OAE-PINN-Emulator.git
cd OAE-PINN-Emulator

2) Create the environment (uv)

If the repo contains pyproject.toml + uv.lock:

uv sync

Run any command like:

uv run python --version

Recommended workflow

Prefer uv run ... over manually activating a virtualenv. It keeps runs consistent and reproducible.

3) Data + outputs layout

Recommended (kept out of git):

# from the repo root
mkdir -p data outputs
  • data/ → NetCDF datasets (Oceananigans output or downloaded sample)
  • outputs/ → checkpoints, predictions, plots, logs

Do not commit large artifacts

Keep NetCDFs, checkpoints, and generated outputs in data//outputs/ (gitignored).

4) Precomputed demo evaluation (placeholder)

This will become the “one command” path.

# placeholder (will be made exact)
uv run python -m oae_pinn.eval --config configs/demo.yaml

Planned outputs: - a figure/GIF saved under outputs/ - a metrics JSON/CSV

5) Train + evaluate from configs (placeholder)

# TRAIN (placeholder)
uv run python -m oae_pinn.train --config configs/baseline.yaml

# EVAL (placeholder)
uv run python -m oae_pinn.eval --config configs/baseline.yaml

To be finalized

These module names and config paths will be updated to match the final repo structure and CLI.