def __init__(self,
device,
batch_size,
iterator,
nfft,
window_length,
window_step,
window=None,
center=None,
num_threads=1,
device_id=0):
super(SpectrogramPipeline, self).__init__(batch_size, num_threads,
device_id)
self.device = device
self.iterator = iterator
self.inputs = ops.ExternalSource()
window_fn = window(
window_length).tolist() if window is not None else None
self.fft = ops.Spectrogram(device=self.device,
nfft=nfft,
window_length=window_length,
window_step=window_step,
window_fn=window_fn,
center_windows=center,
power=2)
# randomly insert extra axis (channels?)
self.r = np.random.randint(-1, 2)
def __init__(self, device, batch_size, nfft, window_length, window_step,
num_threads=1, device_id=0):
super(AudioSpectrogramPipeline, self).__init__(batch_size, num_threads, device_id)
self.input = ops.FileReader(device="cpu", file_root=audio_files)
self.decode = ops.AudioDecoder(device="cpu", dtype=types.FLOAT, downmix=True)
self.fft = ops.Spectrogram(device=device,
nfft=nfft,
window_length=window_length,
window_step=window_step,
power=2)
def __init__(self, device, batch_size, iterator, nfft, window_length, window_step,
num_threads=1, device_id=0):
super(SpectrogramPipeline, self).__init__(batch_size, num_threads, device_id)
self.device = device
self.iterator = iterator
self.inputs = ops.ExternalSource()
self.fft = ops.Spectrogram(device = self.device,
nfft = nfft,
window_length = window_length,
window_step = window_step,
power = 2)
def __init__(self,
device_id,
n_devices,
file_root,
file_list,
batch_size,
sample_rate=16000,
window_size=.02,
window_stride=.01,
nfeatures=64,
nfft=512,
frame_splicing_factor=3,
silence_threshold=-80,
dither=.00001,
preemph_coeff=.97,
lowfreq=0.0,
highfreq=0.0,
num_threads=1):
super().__init__(batch_size, num_threads, device_id, seed=42)
self.dither = dither
self.frame_splicing_factor = frame_splicing_factor
self.read = ops.readers.File(file_root=file_root, file_list=file_list, device="cpu",
shard_id=device_id, num_shards=n_devices)
self.decode = ops.AudioDecoder(device="cpu", dtype=types.FLOAT, downmix=True)
self.normal_distribution = ops.random.Normal(device="cpu")
self.preemph = ops.PreemphasisFilter(preemph_coeff=preemph_coeff)
self.spectrogram = ops.Spectrogram(device="cpu", nfft=nfft,
window_length=window_size * sample_rate,
window_step=window_stride * sample_rate)
self.mel_fbank = ops.MelFilterBank(device="cpu", sample_rate=sample_rate, nfilter=nfeatures,
normalize=True, freq_low=lowfreq, freq_high=highfreq)
self.log_features = ops.ToDecibels(device="cpu", multiplier=np.log(10), reference=1.0,
cutoff_db=-80)
self.get_shape = ops.Shapes(device="cpu")
self.normalize = ops.Normalize(axes=[0], device="cpu")
self.splicing_transpose = ops.Transpose(device="cpu", perm=[1, 0])
self.splicing_reshape = ops.Reshape(device="cpu", rel_shape=[-1, frame_splicing_factor])
self.splicing_pad = ops.Pad(axes=[0], fill_value=0, align=frame_splicing_factor, shape=[1],
device="cpu")
self.get_nonsilent_region = ops.NonsilentRegion(device="cpu", cutoff_db=silence_threshold)
self.trim_silence = ops.Slice(device="cpu", axes=[0])
self.to_float = ops.Cast(dtype=types.FLOAT)
def __init__(self,
device,
file_name,
batch_size,
nfft,
window_length,
window_step,
num_threads=1,
device_id=0):
super(Spectrogram, self).__init__(batch_size, num_threads, device_id)
self.device = device
self.batch_data = []
y, sr = librosa.load(file_name, sr=44100)
for _ in range(batch_size):
self.batch_data.append(np.array(y, dtype=np.float32))
self.external_source = ops.ExternalSource()
self.spectrogram = ops.Spectrogram(device=self.device,
nfft=nfft,
window_length=window_length,
window_step=window_step)
def __init__(self,
device,
batch_size,
nfft,
window_length,
window_step,
center,
layout="ft",
num_threads=1,
device_id=0):
super(AudioSpectrogramPipeline, self).__init__(batch_size, num_threads,
device_id)
self.input = ops.readers.File(device="cpu", files=audio_files)
self.decode = ops.decoders.Audio(device="cpu",
dtype=types.FLOAT,
downmix=True)
self.fft = ops.Spectrogram(device=device,
nfft=nfft,
window_length=window_length,
window_step=window_step,
power=2,
center_windows=center,
layout=layout)
def __init__(
self,
*,
train_pipeline:
bool, # True if train pipeline, False if validation pipeline
device_id,
num_threads,
batch_size,
file_root: str,
file_list: str,
sample_rate,
discrete_resample_range: bool,
resample_range: list,
window_size,
window_stride,
nfeatures,
nfft,
frame_splicing_factor,
dither_coeff,
silence_threshold,
preemph_coeff,
pad_align,
max_duration,
mask_time_num_regions,
mask_time_min,
mask_time_max,
mask_freq_num_regions,
mask_freq_min,
mask_freq_max,
mask_both_num_regions,
mask_both_min_time,
mask_both_max_time,
mask_both_min_freq,
mask_both_max_freq,
preprocessing_device="gpu"):
super().__init__(batch_size, num_threads, device_id)
self._dali_init_log(locals())
if torch.distributed.is_initialized():
shard_id = torch.distributed.get_rank()
n_shards = torch.distributed.get_world_size()
else:
shard_id = 0
n_shards = 1
self.preprocessing_device = preprocessing_device.lower()
assert self.preprocessing_device == "cpu" or self.preprocessing_device == "gpu", \
"Incorrect preprocessing device. Please choose either 'cpu' or 'gpu'"
self.frame_splicing_factor = frame_splicing_factor
assert frame_splicing_factor == 1, "DALI doesn't support frame splicing operation"
self.resample_range = resample_range
self.discrete_resample_range = discrete_resample_range
self.train = train_pipeline
self.sample_rate = sample_rate
self.dither_coeff = dither_coeff
self.nfeatures = nfeatures
self.max_duration = max_duration
self.mask_params = {
'time_num_regions': mask_time_num_regions,
'time_min': mask_time_min,
'time_max': mask_time_max,
'freq_num_regions': mask_freq_num_regions,
'freq_min': mask_freq_min,
'freq_max': mask_freq_max,
'both_num_regions': mask_both_num_regions,
'both_min_time': mask_both_min_time,
'both_max_time': mask_both_max_time,
'both_min_freq': mask_both_min_freq,
'both_max_freq': mask_both_max_freq,
}
self.do_remove_silence = True if silence_threshold is not None else False
self.read = ops.FileReader(device="cpu",
file_root=file_root,
file_list=file_list,
shard_id=shard_id,
num_shards=n_shards,
shuffle_after_epoch=train_pipeline)
# TODO change ExternalSource to Uniform for new DALI release
if discrete_resample_range and resample_range is not None:
self.speed_perturbation_coeffs = ops.ExternalSource(
device="cpu",
cycle=True,
source=self._discrete_resample_coeffs_generator)
elif resample_range is not None:
self.speed_perturbation_coeffs = random.Uniform(
device="cpu", range=resample_range)
else:
self.speed_perturbation_coeffs = None
self.decode = ops.AudioDecoder(
device="cpu",
sample_rate=self.sample_rate if resample_range is None else None,
dtype=types.FLOAT,
downmix=True)
self.normal_distribution = random.Normal(device=preprocessing_device)
self.preemph = ops.PreemphasisFilter(device=preprocessing_device,
preemph_coeff=preemph_coeff)
self.spectrogram = ops.Spectrogram(
device=preprocessing_device,
nfft=nfft,
window_length=window_size * sample_rate,
window_step=window_stride * sample_rate)
self.mel_fbank = ops.MelFilterBank(device=preprocessing_device,
sample_rate=sample_rate,
nfilter=self.nfeatures,
normalize=True)
self.log_features = ops.ToDecibels(device=preprocessing_device,
multiplier=np.log(10),
reference=1.0,
cutoff_db=math.log(1e-20))
self.get_shape = ops.Shapes(device=preprocessing_device)
self.normalize = ops.Normalize(device=preprocessing_device, axes=[1])
self.pad = ops.Pad(device=preprocessing_device,
axes=[1],
fill_value=0,
align=pad_align)
# Silence trimming
self.get_nonsilent_region = ops.NonsilentRegion(
device="cpu", cutoff_db=silence_threshold)
self.trim_silence = ops.Slice(device="cpu",
normalized_anchor=False,
normalized_shape=False,
axes=[0])
self.to_float = ops.Cast(device="cpu", dtype=types.FLOAT)
# Spectrogram masking
self.spectrogram_cutouts = ops.ExternalSource(
source=self._cutouts_generator, num_outputs=2, cycle=True)
self.mask_spectrogram = ops.Erase(device=preprocessing_device,
axes=[0, 1],
fill_value=0,
normalized_anchor=True)
def __init__(self,
device_id,
num_threads,
resample_range: list,
sample_rate=16000,
window_size=0.02,
window_stride=0.01,
window="hann",
normalize="per_feature",
n_fft=None,
preemph=0.97,
nfilt=64,
lowfreq=0,
highfreq=0,
log=True,
dither=constant,
pad_to=8,
max_duration=15.0,
frame_splicing=3,
batch_size=1,
total_samples=16,
audio_fp16_input=True,
device='gpu'):
super().__init__(batch_size, num_threads, device_id,
exec_async=True, exec_pipelined=True, seed =12, prefetch_queue_depth=1)
self._dali_init_log(locals())
if torch.distributed.is_initialized():
shard_id = torch.distributed.get_rank()
n_shards = torch.distributed.get_world_size()
else:
shard_id = 0
n_shards = 1
torch_windows = {
'hann': torch.hann_window,
'hamming': torch.hamming_window,
'blackman': torch.blackman_window,
'bartlett': torch.bartlett_window,
'none': None,
}
self.audio_fp16_input=audio_fp16_input
self.total_samples = total_samples
self.win_length = int(sample_rate * window_size) # frame size
self.hop_length = int(sample_rate * window_stride)
self.n_fft = n_fft or 2 ** math.ceil(math.log2(self.win_length))
self.normalize = normalize
self.log = log
self.dither = dither
self.frame_splicing = frame_splicing
self.nfilt = nfilt
self.preemph = preemph
self.pad_to = pad_to
self.highfreq = highfreq or sample_rate / 2
window_fn = torch_windows.get(window, None)
window_tensor = window_fn(self.win_length,
periodic=False) if window_fn else None
self.sample_rate = sample_rate
self.window_size = window_size
self.window_stride = window_stride
self.window = window_tensor
self.lowfreq = lowfreq
self.log = log
self.device = device
win_unpadded = self.window.tolist()
win_padded = win_unpadded + [0] * (self.n_fft - len(win_unpadded))
print("self.n_fft = {}".format(self.n_fft))
print("self.hop_length = {}".format(self.hop_length))
print("self.win_length = {}".format(self.win_length))
print("self.window_tensor = {}".format(self.window))
print("self.sample_rate = {}".format(self.sample_rate))
print("self.window_size = {}".format(self.window_size))
print("self.window_stride = {}".format(self.window_stride))
print("self.lowfreq = {}".format(self.lowfreq))
print("self.device = {}".format(self.device))
self.extsrc = ops.ExternalSource(name="INPUT_0", device=self.device, no_copy=True)
self.preemph = ops.PreemphasisFilter(preemph_coeff=preemph, device=self.device)
self.spectrogram = ops.Spectrogram(device=self.device,
nfft=self.n_fft,
center_windows=True,
window_fn=win_padded,
window_length=len(win_padded),
window_step=self.hop_length
)
self.mel_fbank = ops.MelFilterBank(device=self.device,
sample_rate=self.sample_rate,
nfilter=self.nfilt,
freq_high=self.highfreq,
freq_low=self.lowfreq,
normalize=normalize
)
self.log_features = ops.ToDecibels(device=self.device, multiplier=np.log(10), reference=1.0,
cutoff_db=math.log(1e-20))
self.get_shape = ops.Shapes(device=self.device)
self.normalize = ops.Normalize(axes=[0], device=self.device, ddof=1)
self.pad = ops.Pad(axes=[0,1], fill_value=0, shape=[502,240], device=self.device)
# Frame splicing
self.splicing_transpose = ops.Transpose(device=self.device, perm=[1, 0])
self.splicing_reshape = ops.Reshape(device=self.device, rel_shape=[-1, self.frame_splicing])
self.splicing_pad = ops.Pad(axes=[0], fill_value=0, align=self.frame_splicing, shape=[1], device=self.device)
self.to_float16 = ops.Cast(dtype=types.FLOAT16, device=self.device)
self.to_float32 = ops.Cast(dtype=types.FLOAT, device=self.device)
self.samples_done = 0
def __init__(self,
*,
pipeline_type,
device_id,
num_threads,
batch_size,
file_root: str,
sampler,
sample_rate,
resample_range: list,
window_size,
window_stride,
nfeatures,
nfft,
dither_coeff,
silence_threshold,
preemph_coeff,
max_duration,
preprocessing_device="gpu"):
super().__init__(batch_size, num_threads, device_id)
self._dali_init_log(locals())
if torch.distributed.is_initialized():
shard_id = torch.distributed.get_rank()
n_shards = torch.distributed.get_world_size()
else:
shard_id = 0
n_shards = 1
self.preprocessing_device = preprocessing_device.lower()
assert self.preprocessing_device == "cpu" or self.preprocessing_device == "gpu", \
"Incorrect preprocessing device. Please choose either 'cpu' or 'gpu'"
self.resample_range = resample_range
train_pipeline = pipeline_type == 'train'
self.train = train_pipeline
self.sample_rate = sample_rate
self.dither_coeff = dither_coeff
self.nfeatures = nfeatures
self.max_duration = max_duration
self.do_remove_silence = True if silence_threshold is not None else False
shuffle = train_pipeline and not sampler.is_sampler_random()
self.read = ops.FileReader(name="Reader",
pad_last_batch=(pipeline_type == 'val'),
device="cpu",
file_root=file_root,
file_list=sampler.get_file_list_path(),
shard_id=shard_id,
num_shards=n_shards,
shuffle_after_epoch=shuffle)
# TODO change ExternalSource to Uniform for new DALI release
if resample_range is not None:
self.speed_perturbation_coeffs = ops.Uniform(device="cpu",
range=resample_range)
else:
self.speed_perturbation_coeffs = None
self.decode = ops.AudioDecoder(
device="cpu",
sample_rate=self.sample_rate if resample_range is None else None,
dtype=types.FLOAT,
downmix=True)
self.normal_distribution = ops.NormalDistribution(
device=preprocessing_device)
self.preemph = ops.PreemphasisFilter(device=preprocessing_device,
preemph_coeff=preemph_coeff)
self.spectrogram = ops.Spectrogram(
device=preprocessing_device,
nfft=nfft,
window_length=window_size * sample_rate,
window_step=window_stride * sample_rate)
self.mel_fbank = ops.MelFilterBank(device=preprocessing_device,
sample_rate=sample_rate,
nfilter=self.nfeatures,
normalize=True)
self.log_features = ops.ToDecibels(device=preprocessing_device,
multiplier=np.log(10),
reference=1.0,
cutoff_db=math.log(1e-20))
self.get_shape = ops.Shapes(device=preprocessing_device)
self.normalize = ops.Normalize(device=preprocessing_device, axes=[1])
self.pad = ops.Pad(device=preprocessing_device, fill_value=0)
# Silence trimming
self.get_nonsilent_region = ops.NonsilentRegion(
device="cpu", cutoff_db=silence_threshold)
self.trim_silence = ops.Slice(device="cpu",
normalized_anchor=False,
normalized_shape=False,
axes=[0])
self.to_float = ops.Cast(device="cpu", dtype=types.FLOAT)
