Zelk Oss Ocean Flow Train Model

What this skill does

This skill is responsible for:

• Running the training entrypoint scripts/train.py from the repository root.
• Ensuring the run happens inside the {{cookiecutter.project_slug}} conda environment.
• Monitoring the training subprocess and periodically reporting whether it is still running.

It does not try to interpret or modify model configs; it simply runs the training command the user provides (or a sensible default) and watches it to completion.

When to use

Use this skill when the user wants to:

• Run model training for an experiment (e.g., python scripts/train.py model=era5 ...).
• Start a new training run and see live progress until completion.
• Confirm a given training command completes successfully (including early failures).

How to run

1. Ensure you are at the repository root.
2. Run training using the {{cookiecutter.project_slug}} conda environment.

The canonical command is:

cd code
conda run -n {{cookiecutter.project_slug}} --no-capture-output python scripts/train.py <hydra overrides...>

Note: conda run avoids needing to source activate and works in non-interactive shells.

Monitoring progress

To satisfy the "repeatedly check if… still running" requirement, run the training command under a small monitor loop that prints a heartbeat every few seconds while the process is alive.

This skill reuses the run-python skill’s monitor helper located at ./.claude/skills/run-python/scripts/monitor_python.py.

Example usage (from the repo root):

python .claude/skills/run-python/scripts/monitor_python.py \
  --env {{cookiecutter.project_slug}} \
  --cwd code \
  --check-interval 10 \
  -- python scripts/train.py model=shallow_water +experiments=shallow_water/det_mse n_epochs=1 device=cpu

This will:

• activate the {{cookiecutter.project_slug}} environment via conda run
• start the training process in ``
• print a heartbeat line every 10 seconds until training exits
• propagate the training process exit code

Example (shallow water smoke test)

A minimal smoke test for the shallow water model (fast, single epoch):

python .claude/skills/run-python/scripts/monitor_python.py \
  --env {{cookiecutter.project_slug}} \
  --cwd code \
  --check-interval 10 \
  -- python scripts/train.py model=shallow_water +experiments=shallow_water/det_mse n_epochs=1 device=cpu

Tips

• If conda run fails, confirm the {{cookiecutter.project_slug}} environment exists (see environment.yml).
• If training is hanging, inspect the latest log output for exceptions and ensure the dataset files are available.
• You can adjust --check-interval to print status more or less frequently.

Source: zelk-oss/ocean_flow — distributed by TomeVault.