coco_pipe.decoding.persistence#

Resumable decoding-run artifact utilities.

Attributes#

RUN_MANIFEST_FILENAME

Functions#

`resolve_primary_metric_name`(available[, metrics])	Pick the primary metric name from available `<metric>_mean` columns.
`stamp_primary_metric`(record[, metrics])	Add `primary_metric_name` + `primary_metric` to a summary record in place.
`redact_sensitive`(value)	Recursively redact credentials before persisting configs or reports.
`config_hash`(config)	Return a stable short hash for a run configuration.
`data_digest`(*values)	Short SHA-256 over array-likes: dtype, shape and contents (`None` allowed).
`make_serializable`(obj)	Recursively convert NumPy types to JSON-safe Python primitives.
`write_run_manifest`(output_dir, manifest)	Write a unit's manifest in one step, so it is never seen half-written.
`read_run_manifest`(output_dir)	Return a unit's manifest, or None if the unit has not finished.
`completed_for_config`(output_dir, config)	Return whether `output_dir` contains a successful matching run.
`load_completed_result_records`(output_dir, context, *)	Rehydrate aggregate rows for a safely resumed experiment.
`write_run_status`(root, status)	Write exactly one run-level success, partial, or failure marker.

Module Contents#

coco_pipe.decoding.persistence.resolve_primary_metric_name(available, metrics=None)#

Pick the primary metric name from available <metric>_mean columns.

Preference order: balanced_accuracy (the group-robust default), then the first configured metric that has a _mean column, then the first available *_mean column. Returns None when no scalar metric column is present.

Parameters:

available (collections.abc.Sequence[str])
metrics (collections.abc.Sequence[str] | None)

Return type:

str | None

coco_pipe.decoding.persistence.stamp_primary_metric(record, metrics=None)#

Add primary_metric_name + primary_metric to a summary record in place.

Keeps every downstream consumer (leaderboards, head-to-head, reports) on one metric contract instead of re-deriving it. No-op when the record carries no scalar *_mean metric column.

Parameters:

record (dict[str, Any])
metrics (collections.abc.Sequence[str] | None)

Return type:

dict[str, Any]

coco_pipe.decoding.persistence.redact_sensitive(value)#

Recursively redact credentials before persisting configs or reports.

Parameters:: value (Any)
Return type:: Any

coco_pipe.decoding.persistence.config_hash(config)#

Return a stable short hash for a run configuration.

Parameters:: config (collections.abc.Mapping[str, Any])
Return type:: str

coco_pipe.decoding.persistence.data_digest(*values)#

Short SHA-256 over array-likes: dtype, shape and contents (None allowed).

The decoding runner uses it for a unit’s input-data fingerprint (part of the resume hash) and for the sample-set signature paired comparisons must share.