Command-line reference¶

From the repository root:

python run.py [plan|train|analyze|run|verify] [--profile NAME] [--settings-dir DIR]

After installation, nnpd is the equivalent command. Both use the same CLI and public execution functions. Run python run.py --help for the parser’s help text.

Actions¶

Action	Reads	Behavior
`plan`	both settings files	Default. Resolve and validate the sweep and print JSON jobs with training and analysis workload estimates; computes nothing.
`train`	`settings_training.py`	Sample/reuse training data, train/reuse models, and stop at checkpoints. Resets each run’s analysis.
`analyze`	`settings_analysis.py`, plus `output` and `runtime` from `settings_training.py`	Analyze the models of the latest successful training. Never trains. Replaces each run’s metrics and plots.
`run`	both settings files	`train`, then `analyze`.
`verify`	`output` from `settings_training.py`	Deep-check committed artifacts in the output folder.

plan, run and analyze check that the analysis of every model fits the limits in settings_analysis.py. If one does not, plan prints the plan and exits with an error, and run and analyze refuse to start before computing anything. train does not depend on the analysis settings and does not check them.

Options¶

Option	Default	Meaning
`--profile NAME`	`DEFAULT_PROFILE` in `settings_training.py`	Select the preset of that name in both settings files.
`--settings-dir DIR`	The working directory	Folder containing `settings_training.py` and `settings_analysis.py`.
`-h`, `--help`	—	Show help and exit.

The experiment class is set by experiment in the training settings and the analysis class by analysis in the analysis settings. Every successful launch copies the settings files it used byte for byte into the output folder, so --settings-dir outputs/<profile> --profile <profile> runs them again in the same output folder.

Do not expect a flag for every hyperparameter. Ordinary knobs deliberately live in the two settings files. GPU visibility and distributed worker counts are selected through the environment and torchrun; see CUDA.

Python equivalents¶

from nnpd import execute, plan

jobs = plan("smoke")  # computes nothing

# Training/analysis is explicit; planning above does neither.
paths = execute("run", profile="smoke")

Valid execute stages are "train", "analyze", and "run". Both functions accept settings_dir= like --settings-dir. For checksums in Python, use nnpd.core.runner.verify_store(output_root).