1 oracle.pipelines.offline_vin

oracle.pipelines.offline_vin

Generate immutable VIN stores with scene labels or actor-only root evidence.

This module owns composition of raw snippet streaming, Oracle labels, optional EVL features, immutable codecs, split assignment, and shard promotion.

1.1 Classes

Name	Description
VinOfflineWriter	Compose raw actor inputs, optional scene labels, features, and immutable codecs.
VinOfflineWriterConfig	Configure immutable scene-labelled V10 or actor-only V11 preparation.

1.1.1 VinOfflineWriter

oracle.pipelines.offline_vin.VinOfflineWriter(config)

Compose raw actor inputs, optional scene labels, features, and immutable codecs.

1.1.1.1 Methods

Name	Description
run	Build an immutable root, sealing a completed prefix on a safe stop.

1.1.1.1.1 run

oracle.pipelines.offline_vin.VinOfflineWriter.run(should_stop=None)

Build an immutable root, sealing a completed prefix on a safe stop.

should_stop is checked before each snippet. The current snippet finishes once, then prepared rows and their exact index are finalized. Existing Ctrl-C prefix finalization is retained; forced process death leaves an unpublished temporary store for controller reconciliation.

1.1.2 VinOfflineWriterConfig

oracle.pipelines.offline_vin.VinOfflineWriterConfig()

Configure immutable scene-labelled V10 or actor-only V11 preparation.

1.1.2.1 Attributes

Name	Description
target_type	Return the generation pipeline factory target.
paths	Project path resolver.
store	Output store configuration.
dataset	Raw ASE/EFM dataset configuration used to stream snippets.
preparation_mode	Persist V10 scene supervision, or V11 actor roots without scene candidate evaluation.
labeler	Scene labeler configuration; actor-only roots reuse only its reference-frame settings.
backbone	Optional EVL backbone configuration.
include_backbone	Whether to persist actor-visible EVL outputs with manifest provenance.
include_depths	Whether scene-labelled rows persist rendered depths; ignored for actor-only roots.
include_pointclouds	Whether rich diagnostic payloads may include candidate point clouds.
include_diagnostic_payloads	Whether to write rich msgpack diagnostic records alongside numeric blocks.
include_gt_obbs	Whether to persist compact GT OBBs as label/evaluation assets.
include_detected_obbs	Whether to persist actor-visible EVL detected boxes from backbone outputs.
include_trajectory_metadata	Whether to persist trajectory timestamps and gravity.
backbone_numeric_keep_fields	EVL fields written as canonical fixed numeric training blocks; None keeps all supported fields.
backbone_payload_keep_fields	EVL backbone fields written to the optional rich diagnostic payload.
vin_pad_points	Fixed K_store row count for world-frame semidense VIN point tensors.
semidense_max_points	Optional cap on collapsed semidense points before padding.
semidense_include_obs_count	Whether VIN points include observation counts.
max_candidates	Maximum number of candidates stored per sample.
samples_per_shard	Number of samples stored in each immutable shard.
max_samples	Optional cap on the number of processed samples.
train_val_split	Fraction of samples assigned to the validation split.
overwrite	Whether an existing store directory may be replaced.
num_failures_allowed	Maximum number of tolerated sample failures before aborting.
verbosity	Verbosity level for dataset build logging.