def forward(ctx, out_fsa: Fsa, unused_in_fsa_scores: torch.Tensor,
arc_map: _k2.RaggedInt) -> torch.Tensor:
if False:
# TODO(fangjun): this is for debugging only. Can be removed.
expected_scores = _k2.index_and_sum(
unused_in_fsa_scores.contiguous(), arc_map)
assert torch.all(torch.eq(out_fsa.scores, expected_scores))
ctx.save_for_backward(unused_in_fsa_scores)
ctx.arc_map = arc_map
return out_fsa.scores
def forward(ctx, src: torch.Tensor, indexes: k2.RaggedInt) -> torch.Tensor:
'''Index a 1-D tensor with a ragged tensor of indexes, perform
a sum-per-sublist operation, and return the resulting 1-D tensor.
Note:
It supports autograd.
Args:
src:
1-D tensor with dtype torch.float32. For example, it can
be a float tensor attribute of an FSA.
indexes:
A ragged tensor with two axes. For example, it can be
the arc map from :func:`_k2.remove_epsilon`
Returns:
1-D torch.Tensor with dtype being `torch.float32`.
'''
assert src.ndim == 1
assert src.dtype == torch.float32
assert indexes.num_axes() == 2
ctx.save_for_backward(src)
ctx.indexes = indexes
ans = _k2.index_and_sum(src, indexes)
return ans