omniloader.transforms.crop¶
Sequence cropping transforms (random window for train, center for eval).
Both crops take a fixed-length window along the sequence axis. A single shared
offset is applied across every targeted sequence key, so time-aligned features
and per-step targets stay in sync. Keys shorter than the requested window are
padded back to length with their placeholder value.
Classes¶
Crop a random contiguous window of the sequence (training augmentation). |
|
Crop the centered window of the sequence (deterministic; used for eval). |
Module Contents¶
- class omniloader.transforms.crop.RandomCrop(length: int, keys: collections.abc.Sequence[str] | None = None, schema: omniloader.schema.spec.UnifiedSchema | None = None)[source]¶
Bases:
omniloader.transforms.base.TransformCrop a random contiguous window of the sequence (training augmentation).
- Parameters:
length – Window length along the sequence axis.
keys – Sequence keys to crop with a shared offset. When
None, every sequence key in the schema (features and aligned targets) is cropped together (requiresschema).schema – Schema used to resolve
keyswhen it isNone.
- apply(sample: omniloader.transforms.base.Sample, generator: torch.Generator | None) omniloader.transforms.base.Sample[source]¶
Crop all targeted keys at one random offset within the valid region.
- class omniloader.transforms.crop.CenterCrop(length: int, keys: collections.abc.Sequence[str] | None = None, schema: omniloader.schema.spec.UnifiedSchema | None = None)[source]¶
Bases:
omniloader.transforms.base.TransformCrop the centered window of the sequence (deterministic; used for eval).
- Parameters:
length – Window length along the sequence axis.
keys – Sequence keys to crop. When
None, every sequence key in the schema is cropped (requiresschema).schema – Schema used to resolve
keyswhen it isNone.
- apply(sample: omniloader.transforms.base.Sample, generator: torch.Generator | None) omniloader.transforms.base.Sample[source]¶
Crop all targeted keys at the centered offset within the valid region.