def supervision_intervals(self, cuts: CutSet) -> Dict[str, torch.Tensor]:
"""
Returns a dict that specifies the start and end bounds for each supervision,
as a 1-D int tensor, in terms of frames:
.. code-block:
{
"sequence_idx": tensor(shape=(S,)),
"start_frame": tensor(shape=(S,)),
"num_frames": tensor(shape=(S,))
}
Where ``S`` is the total number of supervisions encountered in the :class:`CutSet`.
Note that ``S`` might be different than the number of cuts (``B``).
``sequence_idx`` means the index of the corresponding feature matrix (or cut) in a batch.
"""
start_frames, nums_frames = zip(*(supervision_to_frames(
sup, self.extractor.frame_shift, cut.sampling_rate) for cut in cuts
for sup in cut.supervisions))
sequence_idx = [i for i, c in enumerate(cuts) for s in c.supervisions]
return {
"sequence_idx": torch.tensor(sequence_idx, dtype=torch.int32),
"start_frame": torch.tensor(start_frames, dtype=torch.int32),
"num_frames": torch.tensor(nums_frames, dtype=torch.int32),
}
def supervision_intervals(self, cuts: CutSet) -> Dict[str, torch.Tensor]:
"""
Returns a dict that specifies the start and end bounds for each supervision,
as a 1-D int tensor, in terms of frames.
"""
start_frames, nums_frames = zip(*(supervision_to_frames(
sup, self.extractor.frame_shift, cut.sampling_rate) for cut in cuts
for sup in cut.supervisions))
return {
'start_frame': torch.tensor(start_frames, dtype=torch.int32),
'num_frames': torch.tensor(nums_frames, dtype=torch.int32)
}
def __next__(self) -> Dict[str, Union[torch.Tensor, List[str]]]:
"""
Return a new batch, with the batch size automatically determined using the contraints
of max_frames and max_cuts.
"""
# Collect the cuts that will form a batch, satisfying the criteria of max_cuts and max_frames.
# The returned object is a CutSet that we can keep on modifying (e.g. padding, mixing, etc.)
cuts: CutSet = self._collect_batch()
# Sort the cuts by duration so that the first one determines the batch time dimensions.
cuts = cuts.sort_by_duration(ascending=False)
# Perform the padding (and possibly augmentation at the same time).
if self.aug_cuts is not None:
# Mix in the signal from the augmentation CutSet; use them as padding at the same time.
cuts = cuts.mix(self.aug_cuts,
duration=cuts[0].duration,
snr=self.aug_snr,
mix_prob=self.aug_prob)
else:
# We'll just pad it with low energy values to match the longest Cut's duration in the batch.
cuts = cuts.pad()
# Get a tensor with batched feature matrices, shape (B, T, F)
features = collate_features(cuts)
batch = {
'features':
features,
'supervisions':
default_collate(
[{
'sequence_idx': sequence_idx,
'text': supervision.text,
'start_frame': start_frame,
'num_frames': num_frames
} for sequence_idx, cut in enumerate(cuts)
for supervision, (
start_frame, num_frames) in zip(cut.supervisions, (
supervision_to_frames(s,
cut.frame_shift,
cut.sampling_rate,
max_frames=cut.num_frames)
for s in cut.supervisions))])
}
if self.return_cuts:
batch['supervisions']['cut'] = [
cut for cut in cuts for sup in cut.supervisions
]
return batch
def __getitem__(
self,
cut_ids: List[str]) -> Dict[str, Union[torch.Tensor, List[str]]]:
"""
Return a new batch, with the batch size automatically determined using the contraints
of max_frames and max_cuts.
"""
# Collect the cuts that will form a batch, satisfying the criteria of max_cuts and max_frames.
# The returned object is a CutSet that we can keep on modifying (e.g. padding, mixing, etc.)
cuts = self.cuts.subset(cut_ids=cut_ids)
# Sort the cuts by duration so that the first one determines the batch time dimensions.
cuts = cuts.sort_by_duration(ascending=False)
# Optional transforms.
for tnfm in self.cut_transforms:
cuts = tnfm(cuts)
# Get a tensor with batched feature matrices, shape (B, T, F)
# Collation performs auto-padding, if necessary.
features = collate_features(cuts)
batch = {
'features':
features,
'supervisions':
default_collate(
[{
'sequence_idx': sequence_idx,
'text': supervision.text,
'start_frame': start_frame,
'num_frames': num_frames
} for sequence_idx, cut in enumerate(cuts)
for supervision, (
start_frame, num_frames) in zip(cut.supervisions, (
supervision_to_frames(s,
cut.frame_shift,
cut.sampling_rate,
max_frames=cut.num_frames)
for s in cut.supervisions))])
}
if self.return_cuts:
batch['supervisions']['cut'] = [
cut for cut in cuts for sup in cut.supervisions
]
return batch
