quantlab.tracking.base#
The root classes of experiment tracking: where a run’s records are sent.
A tracker decides where the records of a training or a backtest go: to
Weights & Biases, to MLflow, or nowhere (ADR 0015). It is named in a config
and serialised into config.json by class path, like a portfolio
constructor, so a rebuilt run tracks the same way. The default is the
NullTracker, which sends nothing.
Tracker.start_run is a context manager yielding a TrackingRun: one
record holding the config, step metrics (an epoch or a boosting round) and a
summary. Leaving the block finishes the run, also when the body raises, and
the exception propagates. The config is made JSON-safe here, summarize
flattens nested metrics here, and a missing file is refused here, so every
adapter receives the same values.
Adapters live in quantlab/tracking, one module per tracking library; this
module imports none.
Classes
A tracking run that records nothing. |
|
A tracker that sends nothing anywhere: the default. |
|
Where tracking runs are sent; the root of every tracker. |
|
One record a tracker keeps: a training, a CV fold, a CV summary or a backtest. |
Functions
Flatten nested metrics to |