api

api provides the user-facing entry points that wrap the Runner in a CLI-friendly way and expose train, evaluate, and infer as both Python functions and console scripts.

Console Scripts

Command	Function	Notes
mmtrain	train	Forces config.training = True.
mmevaluate	evaluate	Forces config.training = False.
mminfer	infer	Writes predictions to config.result_path.

Each command parses arguments through Chanfig. The default config file is config.yaml in the current working directory; CLI arguments override individual fields:

mmtrain optim.lr=2e-4 epochs=5 data.root=data/my-task

mminfer defaults result_path to ./result.json and emits a RuntimeWarning if the field is not set explicitly.

Working Directory Contract

Before building the runner, dynamic_import appends the parent of cwd to sys.path and imports os.path.basename(cwd) as a Python module. If config.pretrained matches a sub-directory of cwd, that sub-directory is also appended and imported.

This is how custom datasets, models, and runners register themselves before the runner is built. The working directory must therefore be importable as a Python package (i.e. contain an __init__.py).

Programmatic Usage

The same entry points are usable from Python:

from multimolecule.api import train
from multimolecule.runner import Config

config = Config(
    pretrained="multimolecule/rnafm",
    data={"root": "data/my-rna-task", "sequence_cols": ["sequence"], "label_cols": ["activity"]},
    optim={"lr": 1.0e-4, "weight_decay": 1.0e-2, "pretrained_ratio": 1.0e-2},
    epochs=10,
)
train(config)

train / evaluate / infer always run Config.parse on the provided config, so config.yaml (when present) and CLI overrides are still applied. Pass config=None to start from an empty Config and rely entirely on the YAML file plus CLI arguments.

To bypass the CLI orchestration entirely (no banner, no dl.debug context, no NNI integration), build the runner directly through RUNNERS — see the runner page.

Optional Flags

• debug=true — wrap the run in dl.debug for verbose tracing.
• nni=true — read hyperparameters from nni.get_next_parameter() and merge them into the config; requires the nni package.

Result Aggregation

multimolecule.api.stat walks an experiment root, parses each run’s best.json and trainer.yaml, and produces a CSV summary. Run as a script:

python -m multimolecule.api.stat experiment_root=experiments out_path=result.csv