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,
input_size: int,
scaler: Union[int, float] = 255,
mean: List[float] = [0., 0., 0.],
std: List[float] = [1., 1., 1.],
image_pad_value: Union[int, float] = 0,
labels_pad_value: Union[int, float] = -99,
normalize: bool = True):
"""Initialization
Args:
input_size (int): Target size of image resize
scaler (Union[int,float], optional): The scaling factor applied to the input pixel value. Defaults to 255.
mean (List[float], optional): Mean pixel values for image normalization. Defaults to [0.,0.,0.].
std (List[float], optional): Standard deviation values for image normalization. Defaults to [1.,1.,1.].
image_pad_value (Union[int,float], optional): Values of the color to pad the image to square.. Defaults to 0.
labels_pad_value (Union[int,float], optional): Values used to pad the labels information so it have same dimension. Will be deleted on the dataloader. Defaults to -99.
normalize (bool, optional): Will apply normalization if set to True. Defaults to True.
"""
# By default, CropMirrorNormalize divide each pixel by 255, to make it similar with Pytorch Loader behavior
# in which we can control the scaler, we add additional scaler to reverse the effect
self.normalize = normalize
self.image_normalize = ops.CropMirrorNormalize(
device='gpu',
mean=[value * 255 for value in mean],
std=[value * 255 for value in std],
output_layout='CHW',
image_type=types.DALIImageType.BGR)
self.scaler = ops.Normalize(device='gpu',
scale=float(255 / scaler),
mean=0,
stddev=1)
# Padding and resize to prepare tensor output
self.image_pad = ops.Paste(device='gpu',
fill_value=image_pad_value,
ratio=1,
min_canvas_size=input_size,
paste_x=0,
paste_y=0)
self.labels_pad = ops.Pad(device='cpu',
axes=(0, 1),
fill_value=labels_pad_value)
self.model_input_resize = ops.Resize(
device='gpu',
interp_type=types.DALIInterpType.INTERP_CUBIC,
resize_longer=input_size)
self.peek_shape = ops.Shapes(device='gpu')
def __init__(self, batch_size, num_threads, device_id):
super(COCOPipeline, self).__init__(batch_size,
num_threads,
device_id,
exec_async=False,
exec_pipelined=False,
seed=15)
self.input = ops.COCOReader(file_root=file_root,
annotations_file=annotations_file,
shard_id=device_id,
num_shards=num_gpus,
ratio=True,
ltrb=True)
self.decode = ops.ImageDecoder(device="mixed", output_type=types.RGB)
self.flip = ops.Flip(device="gpu")
self.bbflip = ops.BbFlip(device="cpu", ltrb=True)
self.paste_pos = ops.Uniform(range=(0, 1))
self.paste_ratio = ops.Uniform(range=(1, 2))
self.coin = ops.CoinFlip(probability=0.5)
self.coin2 = ops.CoinFlip(probability=0.5)
self.paste = ops.Paste(device="gpu", fill_value=(32, 64, 128))
self.bbpaste = ops.BBoxPaste(device="cpu", ltrb=True)
self.prospective_crop = ops.RandomBBoxCrop(device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0.1, 0.3, 0.5],
scaling=[0.8, 1.0],
ltrb=True)
self.slice = ops.Slice(device="gpu")
# resize
self.resize = ops.Resize(device="gpu",
interp_type=types.INTERP_LINEAR,
resize_shorter=800,
max_size=1200)
self.shape = ops.Shapes(device="gpu")
# normalize and convert hwc to chw
self.cmnp = ops.CropMirrorNormalize(
device="gpu",
output_dtype=types.FLOAT,
output_layout=types.NCHW,
image_type=types.RGB,
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255])
# padding axes=(0,1) -> hwc, axes=(1,2) -> chw
self.padding = ops.Pad(device="gpu",
fill_value=0,
axes=(1, 2),
shape=(800, 1200))
def __init__(self, size, fill_value):
self.rz_img = ops.Resize(device="gpu", resize_longer=size)
self.rz_ann = ops.Resize(device="gpu",
resize_longer=size,
interp_type=types.DALIInterpType.INTERP_NN)
self.cp_img = ops.Crop(device="gpu",
fill_values=fill_value,
out_of_bounds_policy='pad')
self.cp_ann = ops.Crop(device="gpu",
fill_values=0,
out_of_bounds_policy='pad')
self.size = ops.Constant(fdata=size)
self.pos = ops.Uniform(range=[0, 1])
self.scale = ops.Uniform(range=[0.667, 1.5])
self.shape = ops.Shapes(device="gpu")
def __init__(self, file_root, annotations_file,
batch_size, num_threads, device_id=0, num_gpus=1,
mean=(123.675, 116.28, 103.53), stddev=(1., 1., 1.),
random_shuffle=True):
super(COCOPipeline, self).__init__(batch_size, num_threads, device_id, seed = 15)
self.input = ops.COCOReader(file_root=file_root, annotations_file=annotations_file,
shard_id=device_id, num_shards=num_gpus, ratio=True,
skip_empty=True, prefetch_queue_depth=32, random_shuffle=True)
self.decode = ops.ImageDecoder(device='mixed', output_type=types.BGR)
self.resize = ops.Resize(device='gpu', max_size=1216, resize_shorter=800)
self.flip = ops.CoinFlip(device='cpu')
self.bbox_flip = ops.BbFlip(device='gpu')
self.CMN = ops.CropMirrorNormalize(device='gpu', mean=mean, std=stddev, output_layout='HWC')
self.image_pad = ops.Pad(device='gpu', fill_value=0, axes=(0,1), shape=(1216, 1216))
self.bbox_pad = ops.Pad(device='gpu', fill_value=0, axes=(0,), shape=(100,))
self.label_pad = ops.Pad(device='gpu', fill_value=-1, axes=(0,), shape=(100,))
self.get_shape = ops.Shapes(device='gpu')
self.float_cast = ops.Cast(device='gpu', dtype=types.FLOAT)
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)