hylo-data-video¶
Hylo external data sources: folders of image sequences and video files (no action labels)
Repository hylo-harness, adapters/hylo-data-video. Licence: Apache-2.0. Depends on: hylo-core, pillow>=10, imageio>=2.30.
Registered entry points
| Group | Name | Object |
|---|---|---|
| datasources | video.files |
hylo.datasources.video.sources:VideoFileSource |
| datasources | video.frames |
hylo.datasources.video.sources:FrameFolderSource |
| Module | Summary |
|---|---|
hylo.datasources.video |
|
hylo.datasources.video.sources |
Video data sources (RFC-0004 W1): human or robot videos without action labels |
hylo.datasources.video¶
META
module-attribute
¶
META = {'contracts': {'datasources': 'hylo.datasource/v0.1'}, 'maturity': 'implemented', 'plugins': {'datasources/video.files': {'maturity': 'unverified'}}}
hylo.datasources.video.sources¶
Video data sources (RFC-0004 W1): human or robot videos without action labels.
one directory per episode with numbered image files and an optional episode.json
{"instruction": "...", "task_id": "...", "success": true, "camera": "agentview"}.
VideoFileSource: one video file per episode (mp4 needs the ffmpeg extra), with an optional sidecar
Frames carry images only: no state, no actions. Such episodes can train what needs no actions (success judges,
progress models, video world models) and can be labelled by a calibrated judge (core.instrument.label_run);
action labels would need a labeler (latent actions, hand pose), which is not part of this package.
FrameFolderSource ¶
FrameFolderSource(root: str, task_prefix: str = 'video/', camera: str = 'agentview', every: int = 1, size: int | None = 256, max_episodes: int | None = None, licence: str = 'unknown')
Bases: _Base