omniloader.transforms.mix

MixUp / CutMix as a collate-time wrapper.

Cross-sample mixing must happen at the batch level, so this is a collator that wraps a base collate function rather than a per-sample Transform. It mixes only the input feature tensors and emits the metadata a training loop needs to compute the mixed loss itself — the paired (shuffled) targets <target>_b, the mixing coefficient mixup_lambda and the permutation mixup_index. No loss logic lives here, keeping the library strictly data-side.

  • mode="mixup" — convex blend lam * x + (1 - lam) * x[perm].

  • mode="cutmix" — replace a contiguous time span of each sequence with the permuted sample’s span; mixup_lambda is the retained fraction.

Classes

MixupCollator

Wrap a base collator to apply MixUp/CutMix to feature tensors.

Module Contents

class omniloader.transforms.mix.MixupCollator(base_collate: Any, schema: omniloader.schema.spec.UnifiedSchema, alpha: float = 0.2, mode: str = 'mixup', p: float = 1.0, seed: int = 0)[source]

Wrap a base collator to apply MixUp/CutMix to feature tensors.

Parameters:

  • base_collate – The collator producing the batched dict (e.g. unified_collate() or DynamicCollator).

  • schema – The unified schema, used to pick feature/target keys.

  • alpha – Beta distribution parameter; lam ~ Beta(alpha, alpha).

  • mode – "mixup" (blend) or "cutmix" (span replacement).

  • p – Per-batch probability of applying mixing (otherwise passthrough).

  • seed – Seed for the internal NumPy RNG (deterministic across the run).

Raises:

ValueError – If mode is not "mixup" or "cutmix".

__call__(batch: collections.abc.Sequence[dict[str, Any]]) → dict[str, Any][source]

Collate then mix; adds mixup_lambda, mixup_index and <t>_b.