def __init__(self,
ragged: Union[str, _k2.RaggedFloat, _k2.RaggedShape],
values: Optional[torch.Tensor] = None):
'''Construct an instance of :class:`k2.RaggedFloat`.
Args:
ragged:
It can be one of the following types:
- A string. Example value::
[ [1 2] [] [5 10 20] ]
- An instance of :class:`_k2.RaggedFloat`
- An instance of :class:`_k2.RaggedShape`. In this case, you
have to provide the additional argument `values`.
values:
Required only when `ragged` is an instance of
:class:`_k2.RaggedShape`. It is a 1-D torch.Tensor with dtype
torch.float32.
'''
if isinstance(ragged, str):
ragged = _k2.RaggedFloat(ragged)
elif isinstance(ragged, _k2.RaggedShape):
assert values is not None
ragged = _k2.RaggedFloat(ragged, values)
assert isinstance(ragged, _k2.RaggedFloat)
self.ragged = ragged
self._scores = ragged.values()
def test_create_ragged2(self):
lst = [[7, 9], [12, 13], []]
ragged = k2.create_ragged2(lst)
expected = k2.RaggedInt('[[7 9] [12 13] []]')
self.assertEqual(str(ragged), str(expected))
float_lst = [[1.2], [], [3.4, 5.6, 7.8]]
ragged = k2.create_ragged2(float_lst)
expected = _k2.RaggedFloat('[[1.2] [] [3.4 5.6 7.8]]')
self.assertEqual(str(ragged), str(expected))
def total_scores(self) -> _k2.RaggedFloat:
'''Get total scores of the FSAs in this Nbest.
Note:
Since FSAs in Nbest are just linear FSAs, log-semirng and tropical
semiring produce the same total scores.
Returns:
Return a ragged tensor with two axes [utt][path_scores].
'''
scores = self.fsa.get_tot_scores(use_double_scores=True,
log_semiring=False)
# We use single precision here since we only wrap k2.RaggedFloat.
# If k2.RaggedDouble is wrapped, we can use double precision here.
return _k2.RaggedFloat(self.shape, scores.float())
def test_sum_per_sublist(self):
s = '''
0 1 1 0.
0 1 2 0.
0 1 3 0.
1 2 4 0.
1 2 5 0.
2 3 -1 0.
3
'''
fsa = k2.Fsa.from_str(s)
scores = torch.randn_like(fsa.scores)
fsa.set_scores_stochastic_(scores)
normalized_scores = k2.ragged.sum_per_sublist(
_k2.RaggedFloat(fsa.arcs.shape(), fsa.scores.exp()))
assert normalized_scores.numel() == fsa.arcs.dim0()
assert torch.allclose(normalized_scores[:-1],
torch.ones(normalized_scores.numel() - 1))
# the final state has no leaving arcs
assert normalized_scores[-1].item() == 0
def __init__(self, ragged: Union[str, _k2.RaggedFloat]):
if isinstance(ragged, str):
ragged = _k2.RaggedFloat(ragged)
self.ragged = ragged
self._scores = ragged.values()