def test_pytorch_audio_dataset(self):
featurizer = WaveformFeaturizer.from_config(self.featurizer_config)
ds = AudioDataset(manifest_filepath=self.manifest_filepath, labels=self.labels, featurizer=featurizer,)
for i in range(len(ds)):
if i == 5:
logging.info(ds[i])
def __init__(
self,
manifest_filepath,
durs_dir,
labels,
batch_size,
sample_rate=16000,
int_values=False,
bos_id=None,
eos_id=None,
pad_id=None,
min_duration=0.1,
max_duration=None,
normalize_transcripts=True,
trim_silence=False,
load_audio=True,
drop_last=False,
shuffle=True,
num_workers=0,
):
super().__init__()
# Set up dataset.
self._featurizer = WaveformFeaturizer(sample_rate=sample_rate,
int_values=int_values,
augmentor=None)
dataset_params = {
'manifest_filepath': manifest_filepath,
'labels': labels,
'featurizer': self._featurizer,
'max_duration': max_duration,
'min_duration': min_duration,
'normalize': normalize_transcripts,
'trim': trim_silence,
'bos_id': bos_id,
'eos_id': eos_id,
'load_audio': load_audio,
}
audio_dataset = AudioDataset(**dataset_params)
self._dataset = fastspeech.FastSpeechDataset(audio_dataset, durs_dir)
self._pad_id = pad_id
self.sample_rate = sample_rate
sampler = None
if self._placement == nemo.core.DeviceType.AllGpu:
sampler = torch.utils.data.distributed.DistributedSampler(
self._dataset)
self._dataloader = torch.utils.data.DataLoader(
dataset=self._dataset,
batch_size=batch_size,
collate_fn=self._collate,
drop_last=drop_last,
shuffle=shuffle if sampler is None else False,
sampler=sampler,
num_workers=num_workers,
)
def construct_perturbed_dataset(perturbation):
if perturbation is not None:
# Execute perturbations with 100% probability
prob_perturb = [(1.0, perturbation)]
audio_augmentor = perturb.AudioAugmentor(prob_perturb)
else:
audio_augmentor = None
featurizer = WaveformFeaturizer(
sample_rate=self.featurizer_config['sample_rate'],
int_values=self.featurizer_config['int_values'],
augmentor=audio_augmentor,
)
ds = AudioLabelDataset(manifest_filepath=self.manifest_filepath, labels=self.labels, featurizer=featurizer)
return ds
def test_pytorch_audio_dataset_with_perturbation(self):
perturbations = [
perturb.WhiteNoisePerturbation(min_level=-90, max_level=-46),
perturb.ShiftPerturbation(min_shift_ms=-5.0, max_shift_ms=5.0),
]
# Execute perturbations with 100% probability
prob_perturb = [(1.0, p) for p in perturbations]
audio_augmentor = perturb.AudioAugmentor(prob_perturb)
featurizer = WaveformFeaturizer(
sample_rate=self.featurizer_config['sample_rate'],
int_values=self.featurizer_config['int_values'],
augmentor=audio_augmentor,
)
ds = AudioLabelDataset(manifest_filepath=self.manifest_filepath, labels=self.labels, featurizer=featurizer,)
for i in range(len(ds)):
logging.info(ds[i])
def __init__(
self,
data: str,
durs: str,
labels: List[str],
durs_type: str = 'full-pad',
speakers: str = None,
speaker_table: str = None,
speaker_embs: str = None,
batch_size: int = 32,
sample_rate: int = 16000,
int_values: bool = False,
bos_id: Optional[int] = None,
eos_id: Optional[int] = None,
pad_id: Optional[int] = None,
blank_id: Optional[int] = None,
min_duration: Optional[float] = 0.1,
max_duration: Optional[float] = None,
normalize_transcripts: bool = True,
trim_silence: bool = False,
load_audio: bool = True,
drop_last: bool = False,
shuffle: bool = True,
num_workers: int = 0,
sampler_type: str = 'default',
bd_aug: bool = False,
):
"""Creates TalkNet data iterator.
Args:
data: Path to dataset manifest file.
durs: Path to pickled durations file.
labels: List strings of labels to use.
durs_type: String id of durations type to use.
speakers: Speakers list file.
speaker_table: Speakers ids mapping.
speaker_embs: Speakers embeddings file.
batch_size: Number of sample in batch.
sample_rate: Target sampling rate for data. Audio files will be resampled to sample_rate if it is not
already.
int_values: Bool indicating whether the audio file is saved as int data or float data.
bos_id: Beginning of string symbol id used for seq2seq models.
eos_id: End of string symbol id used for seq2seq models.
pad_id: Token used to pad when collating samples in batches.
blank_id: Int id of blank symbol.
min_duration: All training files which have a duration less than min_duration are dropped.
max_duration: All training files which have a duration more than max_duration are dropped.
normalize_transcripts: Whether to use automatic text cleaning.
trim_silence: Whether to use trim silence from beginning and end of audio signal using
librosa.effects.trim().
load_audio: Controls whether the dataloader loads the audio signal and transcript or just the transcript.
drop_last: See PyTorch DataLoader.
shuffle: See PyTorch DataLoader.
num_workers: See PyTorch DataLoader.
sampler_type: String id of sampler type to use.
bd_aug: True if use augmentation for blanks/durs.
"""
super().__init__()
# Set up dataset.
self._featurizer = WaveformFeaturizer(sample_rate=sample_rate, int_values=int_values, augmentor=None)
dataset_params = {
'manifest_filepath': data,
'labels': labels,
'featurizer': self._featurizer,
'max_duration': max_duration,
'min_duration': min_duration,
'normalize': normalize_transcripts,
'trim': trim_silence,
'bos_id': bos_id,
'eos_id': eos_id,
'load_audio': load_audio,
'add_misc': True,
}
audio_dataset = AudioDataset(**dataset_params)
self._dataset = TalkNetDataset(audio_dataset, durs, durs_type, speakers, speaker_table, speaker_embs)
self._durs_type = durs_type
self._pad_id = pad_id
self._blank_id = blank_id
self._space_id = labels.index(' ')
self._sample_rate = sample_rate
self._load_audio = load_audio
self._bd_aug = bd_aug
sampler = None
if self._placement == nemo.core.DeviceType.AllGpu:
if sampler_type == 'all':
sampler = AllSampler(self._dataset)
elif sampler_type == 'default':
sampler = torch.utils.data.distributed.DistributedSampler(self._dataset) # noqa
elif sampler_type == 'super-smart':
sampler = LengthsAwareSampler(
dataset=self._dataset,
lengths=[e.duration for e in audio_dataset.collection],
batch_size=batch_size,
)
else:
raise ValueError("Invalid sample type.")
self._dataloader = torch.utils.data.DataLoader( # noqa
dataset=self._dataset,
batch_size=batch_size,
collate_fn=self._collate,
drop_last=drop_last,
shuffle=shuffle if sampler is None else False,
sampler=sampler,
num_workers=num_workers,
)