def test_error_raised_when_label_len_lower_greater_than_upper(data,
kwargs) -> None:
"""Ensures ``ValueError`` raised when ``label_len[0] > label_leb[1]``."""
upper = kwargs["label_len"][1]
invalid_lower = data.draw(st.integers(min_value=upper + 1))
kwargs["label_len"] = (invalid_lower, upper)
with pytest.raises(ValueError):
speech_to_text(**kwargs)
def random_speech_to_text(
draw, ) -> st.SearchStrategy[Tuple[SpeechToTextGen, Dict]]:
"""Generates different speech_to_text functions."""
kwargs = draw(random_speech_to_text_kwargs())
return speech_to_text(**kwargs), kwargs
def test_error_raised_when_dtype_invalid(data, kwargs) -> None:
"""Ensures ``ValueError`` raised when ``audio_dtype`` invalid."""
invalid_dtypes = [torch.float16, torch.uint8, torch.int8]
kwargs["audio_dtype"] = data.draw(st.sampled_from(invalid_dtypes))
with pytest.raises(ValueError):
speech_to_text(**kwargs)
def test_error_raised_when_audio_channels_less_than_one(data, kwargs) -> None:
"""Ensures ``ValueError`` raised when ``audio_channels < 1``."""
kwargs["audio_channels"] = data.draw(
st.integers(min_value=-1000, max_value=0))
with pytest.raises(ValueError):
speech_to_text(**kwargs)
def test_error_raised_when_label_len_less_than_zero(data, kwargs) -> None:
"""Ensures ``ValueError`` raised when ``label_len[0] < 0``."""
invalid_lower = data.draw(st.integers(min_value=-1000, max_value=-1))
kwargs["label_len"] = (invalid_lower, kwargs["label_len"][1])
with pytest.raises(ValueError):
speech_to_text(**kwargs)
def test_error_raised_when_audio_ms_less_than_one(data, kwargs) -> None:
"""Ensures ``ValueError`` raised when ``audio_ms[0] <= 0``."""
invalid_lower = data.draw(st.integers(min_value=-1000, max_value=0))
kwargs["audio_ms"] = (invalid_lower, kwargs["audio_ms"][1])
with pytest.raises(ValueError):
speech_to_text(**kwargs)
def build(
dataset: dataset_pb2.Dataset,
transform: Optional[Callable] = None,
target_transform: Optional[Callable] = None,
add_seq_len_to_transforms: bool = False,
download: bool = False,
) -> torch.utils.data.Dataset:
"""Returns a :py:class:`torch.utils.data.Dataset` based on the config.
Args:
dataset: A :py:class:`myrtlespeech.protos.dataset_pb2.Dataset` protobuf
object containing the config for the desired
:py:class:`torch.utils.data.Dataset`.
transform: Transform to pass to the
:py:class:`torch.utils.data.Dataset`.
target_transform: Target transform to pass to the
:py:class:`torch.utils.data.Dataset`.
add_seq_len_to_transforms: If :py:data:`True`, an additional function
is applied after ``transform`` and ``target_transform`` that takes
a value and returns a tuple of ``(value,
torch.tensor(len(value)))``.
download: If :py:data:`True` and dataset does not exist, download it
if possible.
Returns:
A :py:class:`torch.utils.data.Dataset` based on the config.
Example:
>>> from google.protobuf import text_format
>>> dataset_cfg = text_format.Merge('''
... fake_speech_to_text {
... dataset_len: 2;
... audio_ms {
... lower: 10;
... upper: 100;
... }
... label_symbols: "abcde";
... label_len {
... lower: 1;
... upper: 10;
... }
... }
... ''', dataset_pb2.Dataset())
>>> dataset = build(dataset_cfg, add_seq_len_to_transforms=True)
>>> len(dataset)
2
>>> (audio, audio_len), (label, label_len) = dataset[0]
>>> type(audio)
<class 'torch.Tensor'>
>>> bool(audio.size(-1) == audio_len)
True
>>> type(label)
<class 'str'>
>>> bool(len(label) == label_len)
True
"""
supported_dataset = dataset.WhichOneof("supported_datasets")
if add_seq_len_to_transforms:
transform = _add_seq_len(transform, len_fn=lambda x: x.size(-1))
target_transform = _add_seq_len(target_transform, len_fn=len)
if supported_dataset == "fake_speech_to_text":
cfg = dataset.fake_speech_to_text
dataset = FakeDataset(
generator=speech_to_text(
audio_ms=(cfg.audio_ms.lower, cfg.audio_ms.upper),
label_symbols=cfg.label_symbols,
label_len=(cfg.label_len.lower, cfg.label_len.upper),
audio_transform=transform,
label_transform=target_transform,
),
dataset_len=cfg.dataset_len,
)
elif supported_dataset == "librispeech":
cfg = dataset.librispeech
max_duration = cfg.max_secs.value if cfg.HasField("max_secs") else None
dataset = LibriSpeech(
root=cfg.root,
subsets=[
cfg.SUBSET.DESCRIPTOR.values_by_number[subset_idx].name.lower(
).replace("_", "-") for subset_idx in cfg.subset
],
audio_transform=transform,
label_transform=target_transform,
download=download,
max_duration=max_duration,
)
else:
raise ValueError(f"{supported_dataset} not supported")
return dataset