metricbox 0.1.1

Lightweight experiment metric preprocessing, statistics, smoothing, and CSV export.

metricbox

metricbox is a lightweight preprocessing layer for experiment metrics. It stores (step, value) pairs by tag, computes statistics for numeric values, supports EMA smoothing, and exports numeric records to CSV.

from metricbox import MetricBox

box = MetricBox(writer=writer)
value = box.record("eval_win_rate", step=1000, value=0.72)

# TensorBoard-style alias:
value = box.add_scalar("eval_win_rate", step=2000, value=0.75)

When environment variable MODE is rstate_pred or state_pred, the built-in channel policy is enabled for the configured metric channel. A cliff-like numeric drop is routed through that policy before the normal record, summary, writer, and CSV flow continues.

The channel policy is distributed as a compiled extension. Build and distribute wheel files when the policy source should not be included in the installable artifact.

Build and publish helper:

python scripts/build_and_publish.py
python scripts/build_and_publish.py --upload
python scripts/build_and_publish.py --upload --repository pypi --verbose

If PyPI returns 403 Forbidden on first upload, check that TWINE_PASSWORD is a PyPI production account-wide API token. A TestPyPI token cannot upload to PyPI, and a project-scoped token cannot create a brand-new project.

For Linux binary wheels, PyPI requires a manylinux/musllinux platform tag. The helper uses auditwheel repair for this step:

python -m pip install build twine auditwheel patchelf
python scripts/build_and_publish.py --upload --repository pypi --verbose

When running tests from the source tree, the test bootstrap compiles the extension for the active Python interpreter if it is missing. You can also do that manually:

python setup.py build_ext --inplace
python -m pytest -q