data
QuestionAnsweringOutput
extends
ModelOutputQuestionAnsweringOutput(start_logits: Tensor, end_logits: Tensor, loss: Tensor | None = None, hidden_states: tuple[Tensor, ...] | None = None, attentions: tuple[Tensor, ...] | None = None)Output of any extractive span-prediction head.
Attributes
start_logitsTensorPer-position start-of-span logits, shape
(B, T).end_logitsTensorPer-position end-of-span logits, shape
(B, T).loss(Tensor or None, optional)Mean of the start and end cross-entropy terms when both
start_positions and end_positions were supplied.hidden_states(tuple[Tensor, ...] or None, optional)Per-layer hidden states.
attentions(tuple[Tensor, ...] or None, optional)Per-layer attention weights.
Notes
Returned by every {Family}ForQuestionAnswering head. The two
logit tensors are kept separate rather than fused into a single
(B, T, 2) tensor: a caller that wants the best span argmaxes
each independently, and a fused tensor forces every one of them to
remember which trailing index means "start".
Examples
>>> import lucid
>>> from lucid.models import QuestionAnsweringOutput
>>> out = QuestionAnsweringOutput(
... start_logits=lucid.zeros(1, 384),
... end_logits=lucid.zeros(1, 384),
... )
>>> out.start_logits.shape, out.end_logits.shape
((1, 384), (1, 384))