def __init__(self, device, batch_size, num_threads=1, device_id=0):
super(MultichannelPipeline, self).__init__(batch_size, num_threads,
device_id)
self.device = device
self.reader = ops.readers.File(files=multichannel_tiff_files)
decoder_device = 'mixed' if self.device == 'gpu' else 'cpu'
self.decoder = ops.decoders.Image(device=decoder_device,
output_type=types.ANY_DATA)
self.resize = ops.Resize(device=self.device,
resize_y=900,
resize_x=300,
min_filter=types.DALIInterpType.INTERP_LINEAR)
self.crop = ops.Crop(device=self.device,
crop_h=220,
crop_w=224,
crop_pos_x=0.3,
crop_pos_y=0.2)
self.transpose = ops.Transpose(device=self.device,
perm=(1, 0, 2),
transpose_layout=False)
self.cmn = ops.CropMirrorNormalize(device=self.device,
std=255.,
mean=0.,
output_layout="HWC",
dtype=types.FLOAT)
def __init__(self, DATA_PATH, input_height, batch_size, num_threads,
device_id):
super(SimCLRFinetuneTrainDataTransform, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
self.COPIES = 1
self.input_height = input_height
self.input = ops.FileReader(file_root=DATA_PATH,
random_shuffle=True,
seed=12)
self.coin = ops.CoinFlip(probability=0.5)
self.uniform = ops.Uniform(range=[0.7, 1.3]) #-1 to 1
#read image (I think that has to be cpu, do a mixed operation to decode into gpu)
self.decode = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.crop = ops.RandomResizedCrop(size=self.input_height,
minibatch_size=batch_size,
device="gpu")
self.flip = ops.Flip(vertical=self.coin(),
horizontal=self.coin(),
device="gpu")
self.colorjit_gray = ops.ColorTwist(brightness=self.uniform(),
contrast=self.uniform(),
hue=self.uniform(),
saturation=self.uniform(),
device="gpu")
self.blur = ops.GaussianBlur(window_size=int(0.1 * self.input_height),
device="gpu",
dtype=types.FLOAT)
self.swapaxes = ops.Transpose(perm=[2, 0, 1], device="gpu")
self.to_int64 = ops.Cast(dtype=types.INT64, device="gpu")
def __init__(self, params, num_threads, device_id):
super(DaliPipeline, self).__init__(params.batch_size,
num_threads,
device_id,
seed=12)
dii = DaliInputIterator(params, device_id)
self.no_copy = params.no_copy
if self.no_copy:
print("Use Zero Copy ES")
self.source = ops.ExternalSource(device="gpu",
source=dii,
num_outputs=2,
layout=["DHWC", "DHWC"],
no_copy=self.no_copy)
self.do_rotate = True if params.rotate_input == 1 else False
print("Enable Rotation" if self.do_rotate else "Disable Rotation")
self.rng_angle = ops.Uniform(device="cpu", range=[-1.5, 2.5])
self.icast = ops.Cast(device="cpu", dtype=types.INT32)
self.fcast = ops.Cast(device="cpu", dtype=types.FLOAT)
self.rotate1 = ops.Rotate(device="gpu",
axis=(1, 0, 0),
interp_type=types.INTERP_LINEAR)
self.rotate2 = ops.Rotate(device="gpu",
axis=(0, 1, 0),
interp_type=types.INTERP_LINEAR)
self.rotate3 = ops.Rotate(device="gpu",
axis=(0, 0, 1),
interp_type=types.INTERP_LINEAR)
self.transpose = ops.Transpose(device="gpu", perm=[3, 0, 1, 2])
def __init__(self,
DATA_PATH,
input_height,
batch_size,
copies,
stage,
num_threads,
device_id,
seed=1729):
super(SimCLRTransform, self).__init__(batch_size,
num_threads,
device_id,
seed=seed)
#this lets our pytorch compat function find the length of our dataset
self.num_samples = len(ImageFolder(DATA_PATH))
self.copies = copies
self.input_height = input_height
self.stage = stage
self.input = ops.FileReader(file_root=DATA_PATH,
random_shuffle=True,
seed=seed)
self.to_int64 = ops.Cast(dtype=types.INT64, device="gpu")
self.to_int32_cpu = ops.Cast(dtype=types.INT32, device="cpu")
self.coin = ops.random.CoinFlip(probability=0.5)
self.uniform = ops.random.Uniform(range=[0.6, 0.9])
self.blur_amt = ops.random.Uniform(values=[
float(i) for i in range(1, int(0.1 * self.input_height), 2)
])
self.angles = ops.random.Uniform(range=[0, 360])
self.cast = ops.Cast(dtype=types.FLOAT, device='gpu')
self.decode = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.crop = ops.RandomResizedCrop(size=self.input_height,
minibatch_size=batch_size,
random_area=[0.75, 1.0],
device="gpu")
self.resize = ops.Resize(resize_x=self.input_height,
resize_y=self.input_height,
device="gpu")
self.flip = ops.Flip(vertical=self.coin(),
horizontal=self.coin(),
device="gpu")
self.colorjit_gray = ops.ColorTwist(brightness=self.uniform(),
contrast=self.uniform(),
hue=self.uniform(),
saturation=self.uniform(),
device="gpu")
self.blur = ops.GaussianBlur(window_size=self.to_int32_cpu(
self.blur_amt()),
device="gpu")
self.rotate = ops.Rotate(
angle=self.angles(),
keep_size=True,
interp_type=types.DALIInterpType.INTERP_LINEAR,
device="gpu")
self.swapaxes = ops.Transpose(perm=[2, 0, 1], device="gpu")
def __init__(self, params, device_id, files, labels):
super().__init__(params.batch_size,
params.num_gpus * 8,
device_id,
seed=params.seed)
# file_root有坑，并不是文件夹名字就是label，按照文件夹顺序(1, 10, 11, 2, 20, 21, ...)分别给与0，1，2，3，4...标签
self.input = ops.FileReader(files=files,
labels=labels,
random_shuffle=True)
self.decocer = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.resize = ops.Resize(device='gpu', resize_shorter=224)
self.pos_rng_x = ops.random.Uniform(range=(0.0, 1.0))
self.pos_rng_y = ops.random.Uniform(range=(0.0, 1.0))
self.crop = ops.Crop(device='gpu', crop_h=224, crop_w=224)
self.flip = ops.Flip(device='gpu')
self.coinflip = ops.random.CoinFlip(probability=0.5)
self.hsv = ops.Hsv(device='gpu')
self.saturation = ops.random.Uniform(range=(0.8, 1.0))
self.value = ops.random.Uniform(range=(0.8, 1.0))
mean = torch.Tensor(params.mean).unsqueeze(0).unsqueeze(0) * 255
std = torch.Tensor(params.std).unsqueeze(0).unsqueeze(0) * 255
self.normalize = ops.Normalize(axes=[0, 1],
mean=mean,
stddev=std,
device='gpu',
batch=False)
self.transpose = ops.Transpose(device='gpu', perm=[2, 0, 1])
def __init__(self,
batch_size,
num_threads,
device_id,
file_list,
sequence_length,
seg_num,
seg_length,
resize_shorter_scale,
crop_target_size,
is_training=False,
initial_prefetch_size=10,
num_shards=1,
shard_id=0,
dali_mean=0.,
dali_std=1.0):
super(VideoPipe, self).__init__(batch_size, num_threads, device_id)
self.input = ops.VideoReader(device="gpu",
file_list=file_list,
sequence_length=sequence_length,
seg_num=seg_num,
seg_length=seg_length,
is_training=is_training,
num_shards=num_shards,
shard_id=shard_id,
random_shuffle=is_training,
initial_fill=initial_prefetch_size)
# the sequece data read by ops.VideoReader is of shape [F, H, W, C]
# Because the ops.Resize does not support sequence data,
# it will be transposed into [H, W, F, C],
# then reshaped to [H, W, FC], and then resized like a 2-D image.
self.transpose = ops.Transpose(device="gpu", perm=[1, 2, 0, 3])
self.reshape = ops.Reshape(device="gpu",
rel_shape=[1.0, 1.0, -1],
layout='HWC')
self.resize = ops.Resize(device="gpu",
resize_shorter=resize_shorter_scale)
# crops and mirror are applied by ops.CropMirrorNormalize.
# Normalization will be implemented in paddle due to the difficulty of dimension broadcast,
# It is not sure whether dimension broadcast can be implemented correctly by dali, just take the Paddle Op instead.
self.pos_rng_x = ops.Uniform(range=(0.0, 1.0))
self.pos_rng_y = ops.Uniform(range=(0.0, 1.0))
self.mirror_generator = ops.Uniform(range=(0.0, 1.0))
self.cast_mirror = ops.Cast(dtype=types.DALIDataType.INT32)
self.crop_mirror_norm = ops.CropMirrorNormalize(
device="gpu",
crop=[crop_target_size, crop_target_size],
mean=dali_mean,
std=dali_std)
self.reshape_back = ops.Reshape(device="gpu",
shape=[
seg_num, seg_length * 3,
crop_target_size, crop_target_size
],
layout='FCHW')
self.cast_label = ops.Cast(device="gpu",
dtype=types.DALIDataType.INT64)
def __init__(self, device, batch_size, layout, iterator, num_threads=1, device_id=0,
permutation = (1, 0, 2), transpose_layout=False, out_layout_arg=None):
super(TransposePipeline, self).__init__(batch_size,
num_threads,
device_id)
self.device = device
self.layout = layout
self.iterator = iterator
self.inputs = ops.ExternalSource()
if out_layout_arg:
self.transpose = ops.Transpose(device = self.device,
perm = permutation,
transpose_layout = transpose_layout,
output_layout = out_layout_arg)
else:
self.transpose = ops.Transpose(device = self.device,
perm = permutation,
transpose_layout = transpose_layout)
def __init__(self, device, batch_size, layout, iterator, num_threads=1, device_id=0):
super(TransposePipeline, self).__init__(batch_size,
num_threads,
device_id)
self.device = device
self.layout = layout
self.iterator = iterator
self.inputs = ops.ExternalSource()
self.transpose = ops.Transpose(device = self.device,
perm = (1, 0, 2))
def __init__(self, image_dir, batch_size, num_threads, device_id):
super(SimplePipeline, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
self.input = ops.FileReader(file_root=image_dir)
# instead of path to file directory file with pairs image_name image_label_value can be provided
# self.input = ops.FileReader(file_root = image_dir, file_list = image_dir + '/file_list.txt')
self.decode = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.transpose = ops.Transpose(device='gpu', perm=[2, 0, 1])
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, batch_size, sequence_length, num_threads, device_id,
files):
super(VideoReaderPipeline, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
self.reader = ops.VideoReader(device="gpu",
filenames=files,
sequence_length=sequence_length,
normalized=False,
image_type=types.RGB,
dtype=types.UINT8)
self.transpose = ops.Transpose(device="gpu", perm=[0, 3, 1, 2])
def __init__(self, params, num_threads, device_id):
super(DaliPipeline, self).__init__(params.batch_size,
num_threads,
device_id,
seed=12)
with h5py.File(params.data_path, 'r') as f:
# load hydro and clean up
Hydro = f['Hydro'][...]
self.Hydro = types.Constant(Hydro,
shape=Hydro.shape,
layout="DHWC",
device="cpu")
del Hydro
# load nbody and clean up
Nbody = f['Nbody'][...]
self.Nbody = types.Constant(Nbody,
shape=Nbody.shape,
layout="DHWC",
device="cpu")
del Nbody
#self.ndummy = np.zeros((20, 20, 20, 4), dtype=np.float32)
#self.hdummy = np.zeros((20, 20, 20, 5), dtype=np.float32)
#self.Nbody = types.Constant(self.ndummy, shape = self.ndummy.shape, layout = "DHWC", device="cpu")
#self.Hydro = types.Constant(self.hdummy, shape = self.hdummy.shape, layout = "DHWC", device="cpu")
#self.Nbody = ops.Constant(fdata = self.ndummy.flatten().tolist(), shape = self.ndummy.shape, layout = "DHWC", device = "cpu")
#self.Hydro = ops.Constant(fdata = self.hdummy.flatten().tolist(), shape = self.hdummy.shape, layout = "DHWC", device = "cpu")
self.do_rotate = True if params.rotate_input == 1 else False
print("Enable Rotation" if self.do_rotate else "Disable Rotation")
self.rng_angle = ops.Uniform(device="cpu", range=[-1.5, 2.5])
self.rng_pos = ops.Uniform(device="cpu", range=[0., 1.])
self.icast = ops.Cast(device="cpu", dtype=types.INT32)
self.fcast = ops.Cast(device="cpu", dtype=types.FLOAT)
self.crop = ops.Crop(device="cpu",
crop_d=params.data_size,
crop_h=params.data_size,
crop_w=params.data_size)
self.rotate1 = ops.Rotate(device="gpu",
axis=(1, 0, 0),
interp_type=types.INTERP_LINEAR)
self.rotate2 = ops.Rotate(device="gpu",
axis=(0, 1, 0),
interp_type=types.INTERP_LINEAR)
self.rotate3 = ops.Rotate(device="gpu",
axis=(0, 0, 1),
interp_type=types.INTERP_LINEAR)
self.transpose = ops.Transpose(device="gpu", perm=[3, 0, 1, 2])
def __init__(self, batch_size, num_threads, device_id):
super(TransposePipe, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
self.input = ops.CaffeReader(path=caffe_db_folder,
shard_id=device_id,
num_shards=1)
self.decode = ops.nvJPEGDecoder(device="mixed",
output_type=types.RGB)
self.crop = ops.Crop(device="gpu",
crop=(224, 224),
image_type=types.RGB)
self.transpose = ops.Transpose(device="gpu", perm=[2, 0, 1])
def __init__(self, batch_size, num_threads, device_id, tfrecords,
idx_paths):
super(ResnetPipeline, self).__init__(batch_size, num_threads,
device_id)
# Transformation operations below.
# From https://docs.nvidia.com/deeplearning/sdk/dali-developer-guide/docs/supported_ops.html
self.input = ops.TFRecordReader(
path=tfrecords,
index_path=idx_paths,
features={
"image/encoded": tfrec.FixedLenFeature([], tfrec.string, ""),
"image/class/label": tfrec.FixedLenFeature([1], tfrec.float32,
0.0),
"image/class/text": tfrec.FixedLenFeature([], tfrec.string,
""),
"image/object/bbox/xmin":
tfrec.VarLenFeature(tfrec.float32, 0.0),
"image/object/bbox/ymin":
tfrec.VarLenFeature(tfrec.float32, 0.0),
"image/object/bbox/xmax":
tfrec.VarLenFeature(tfrec.float32, 0.0),
"image/object/bbox/ymax":
tfrec.VarLenFeature(tfrec.float32, 0.0)
})
self.decode = ops.nvJPEGDecoder(device="mixed",
cache_debug=True,
output_type=types.RGB)
self.resize = ops.Resize(device="gpu",
image_type=types.RGB,
interp_type=types.INTERP_LINEAR,
resize_shorter=256.)
self.cmn = ops.CropMirrorNormalize(device="gpu",
output_dtype=types.FLOAT,
crop=(224, 224),
image_type=types.RGB,
mean=[0., 0., 0.],
std=[1., 1., 1])
self.uniform = ops.Uniform(range=(0.0, 1.0))
self.transpose = ops.Transpose(device="gpu", perm=[0, 3, 1, 2])
self.cast = ops.Cast(device="gpu", dtype=types.INT32)
self.iter = 0
def __init__(self, batch_size, sequence_length, num_threads, device_id,
files, crop_size):
super(VideoReaderPipeline, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
self.reader = ops.VideoReader(device="gpu",
filenames=files,
sequence_length=sequence_length,
normalized=False,
random_shuffle=True,
image_type=types.RGB,
dtype=types.UINT8,
initial_fill=16)
# self.crop = ops.Crop(device="gpu", crop=crop_size, output_dtype=types.FLOAT)
self.uniform = ops.Uniform(range=(0.0, 1.0))
self.transpose = ops.Transpose(device="gpu", perm=[3, 0, 1, 2])
def __init__(self, files, labels):
super().__init__(256, 8, 0, seed=42)
self.input = ops.FileReader(files=files,
labels=labels,
random_shuffle=False)
self.decocer = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.resize = ops.Resize(device='gpu', resize_shorter=224)
self.crop = ops.Crop(device='gpu', crop_h=224, crop_w=224)
mean = torch.Tensor([0.485, 0.456, 0.406
]).unsqueeze(0).unsqueeze(0) * 255
std = torch.Tensor([0.229, 0.224, 0.225
]).unsqueeze(0).unsqueeze(0) * 255
self.normalize = ops.Normalize(axes=[0, 1],
mean=mean,
stddev=std,
device='gpu',
batch=False)
self.transpose = ops.Transpose(device='gpu', perm=[2, 0, 1])
def __init__(self, params, device_id, files, labels):
super().__init__(params.batch_size,
params.num_gpus * 8,
device_id,
seed=params.seed)
self.input = ops.FileReader(files=files,
labels=labels,
random_shuffle=False)
self.decocer = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.resize = ops.Resize(device='gpu', resize_shorter=224)
self.crop = ops.Crop(device='gpu', crop_h=224, crop_w=224)
mean = torch.Tensor(params.mean).unsqueeze(0).unsqueeze(0) * 255
std = torch.Tensor(params.std).unsqueeze(0).unsqueeze(0) * 255
self.normalize = ops.Normalize(axes=[0, 1],
mean=mean,
stddev=std,
device='gpu',
batch=False)
self.transpose = ops.Transpose(device='gpu', perm=[2, 0, 1])
def __init__(self, device, batch_size, iterator, nfilter, sample_rate, freq_low, freq_high,
normalize, mel_formula, layout='ft', num_threads=1, device_id=0, func=mel_fbank_func):
super(MelFilterBankPythonPipeline, self).__init__(
batch_size, num_threads, device_id,
seed=12345, exec_async=False, exec_pipelined=False)
self.device = "cpu"
self.iterator = iterator
self.inputs = ops.ExternalSource()
function = partial(func, nfilter, sample_rate, freq_low, freq_high, normalize, mel_formula)
self.mel_fbank = ops.PythonFunction(function=function)
self.layout=layout
self.freq_major = layout.find('f') != len(layout) - 1
self.need_transpose = not self.freq_major and len(layout) > 1
if self.need_transpose:
perm = [i for i in range(len(layout))]
f = layout.find('f')
perm[f] = len(layout) - 2
perm[-2] = f
self.transpose = ops.Transpose(perm=perm)
def __init__(self, params, num_threads, device_id):
super(DaliPipeline, self).__init__(params.batch_size,
num_threads,
device_id,
seed=12)
dii = DaliInputIterator(params, device_id)
dpi = DaliParameterIterator(params)
self.no_copy = params.no_copy
if self.no_copy:
print("Use Zero Copy ES")
self.source = ops.ExternalSource(device="gpu",
source=dii,
num_outputs=2,
layout=["DHWC", "DHWC"],
no_copy=self.no_copy)
self.params = ops.ExternalSource(device="cpu",
source=dpi,
num_outputs=2)
self.do_rotate = True if params.rotate_input == 1 else False
print("Enable Rotation" if self.do_rotate else "Disable Rotation")
self.rotate = ops.Rotate(device="gpu", interp_type=types.INTERP_LINEAR)
self.transpose = ops.Transpose(device="gpu", perm=[3, 0, 1, 2])
def __init__(self, DATA_PATH, input_height, batch_size, num_threads,
device_id):
super(SimCLRFinetuneValDataTransform, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
self.COPIES = 1
self.input_height = input_height
self.input = ops.FileReader(file_root=DATA_PATH,
random_shuffle=True,
seed=12)
self.decode = ops.ImageDecoder(device='mixed', output_type=types.RGB)
self.crop = ops.RandomResizedCrop(size=self.input_height,
random_area=1,
random_aspect_ratio=1,
minibatch_size=batch_size,
device="gpu",
dtype=types.FLOAT)
self.swapaxes = ops.Transpose(perm=[2, 0, 1], device="gpu")
self.to_int64 = ops.Cast(dtype=types.INT64, device="gpu")
def __init__(self,
device,
batch_size,
layout,
iterator,
num_threads=1,
device_id=0,
tested_operator=None):
super(MultichannelSynthPipeline,
self).__init__(batch_size, num_threads, device_id)
self.device = device
self.layout = layout
self.iterator = iterator
self.inputs = ops.ExternalSource()
self.tested_operator = tested_operator
if self.tested_operator == 'resize' or not self.tested_operator:
self.resize = ops.Resize(
device=self.device,
resize_y=900,
resize_x=300,
min_filter=types.DALIInterpType.INTERP_LINEAR)
if self.tested_operator == 'crop' or not self.tested_operator:
self.crop = ops.Crop(device=self.device,
crop=(220, 224),
crop_pos_x=0.3,
crop_pos_y=0.2,
image_type=types.ANY_DATA)
if self.tested_operator == 'transpose' or not self.tested_operator:
self.transpose = ops.Transpose(device=self.device,
perm=(1, 0, 2),
transpose_layout=False)
if self.tested_operator == 'normalize' or not self.tested_operator:
self.cmn = ops.CropMirrorNormalize(device=self.device,
std=255.,
mean=0.,
output_layout="HWC",
output_dtype=types.FLOAT)
def __init__(self, params, num_threads, device_id):
super(DaliPipeline, self).__init__(params.batch_size,
num_threads,
device_id,
seed=12)
dii = DaliInputIterator(params)
self.no_copy = params.no_copy
if self.no_copy:
print("Use Zero Copy ES")
self.source = ops.ExternalSource(source = dii, num_outputs = 2, layout = ["DHWC", "DHWC"], no_copy = self.no_copy)
self.do_rotate = True if params.rotate_input==1 else False
print("Enable Rotation" if self.do_rotate else "Disable Rotation")
self.rng_angle = ops.Uniform(device = "cpu",
range = [0., 180.])
self.rng_axis = ops.Uniform(device = "cpu",
range = [-1., 1.],
shape=(3))
self.rotate = ops.Rotate(device = "gpu",
interp_type = types.INTERP_LINEAR,
keep_size=True)
self.transpose = ops.Transpose(device = "gpu",
perm=[3,0,1,2])
self.crop = ops.Crop(device = "gpu",
crop = (dii.size, dii.size, dii.size))
def __init__(self, batch_size, sequence_length, num_threads, files, gt_files, \
crop_size, random_shuffle=True, step=-1, device_id=-1, seed=12):
super(VideoReaderPipeline, self).__init__(batch_size,
num_threads,
device_id,
seed=seed)
#Define VideoReader
self.reader = ops.VideoReader(device="gpu", \
filenames=files, \
sequence_length=sequence_length, \
normalized=False, \
random_shuffle=random_shuffle, \
image_type=types.RGB, \
dtype=types.UINT8, \
step=step, \
initial_fill=16,
seed=seed)
self.gt_reader = ops.VideoReader(device="gpu", \
filenames=gt_files, \
sequence_length=sequence_length, \
normalized=False, \
random_shuffle=random_shuffle, \
image_type=types.RGB, \
dtype=types.UINT8, \
step=step, \
initial_fill=16,
seed=seed)
# Define crop and permute operations to apply to every sequence
self.crop = ops.Crop(device="gpu", \
crop=[crop_size, crop_size], \
output_dtype=types.FLOAT)
self.uniform = ops.Uniform(range=(0.0, 1.0)) # used for random crop
self.transpose = ops.Transpose(device="gpu", perm=[3, 0, 1, 2])
def __init__(self,
file_list,
batch_size,
sequence_length,
num_threads,
device_id,
crop_size,
step=-1,
stride=1,
random_shuffle=True):
super(VideoReaderPipeline, self).__init__(batch_size,
num_threads,
device_id,
seed=12)
# Define VideoReader
self.reader = ops.VideoReader(device="gpu",
file_list=file_list,
sequence_length=sequence_length,
normalized=False,
random_shuffle=random_shuffle,
image_type=types.RGB,
dtype=types.UINT8,
step=step,
stride=stride,
initial_fill=16)
# Define crop, mirror and normalisation operations to apply to every sequence
self.crop = ops.Crop(device="gpu",
crop=crop_size,
output_dtype=types.FLOAT)
self.transpose = ops.Transpose(device="gpu", perm=[0, 3, 1,
2]) # [N F C H W]
self.uniform = ops.Uniform(range=(0.2, 1.0)) # used for random crop
def __init__(self,
directory,
supervised: bool = True,
sequence_length: int = 11,
batch_size: int = 1,
num_workers: int = 1,
gpu_id: int = 0,
shuffle: bool = True,
crop_size: tuple = (256, 256),
resize: tuple = None,
brightness: float = 0.25,
contrast: float = 0.1,
mean: list = [0.5, 0.5, 0.5],
std: list = [0.5, 0.5, 0.5],
conv_mode='3d',
image_shape=(256, 256),
validate: bool = False):
super().__init__(batch_size,
num_workers,
gpu_id,
prefetch_queue_depth=1)
self.input = ops.VideoReader(additional_decode_surfaces=1,
channels=3,
device="gpu",
dtype=types.FLOAT,
enable_frame_num=False,
enable_timestamps=False,
file_root=directory,
image_type=types.RGB,
initial_fill=1,
lazy_init=False,
normalized=True,
num_shards=1,
pad_last_batch=False,
prefetch_queue_depth=1,
random_shuffle=shuffle,
sequence_length=sequence_length,
skip_vfr_check=True,
step=-1,
shard_id=0,
stick_to_shard=False,
stride=1)
self.uniform = ops.Uniform(range=(0.0, 1.0))
self.cmn = ops.CropMirrorNormalize(device='gpu',
crop=crop_size,
mean=mean,
std=std,
output_layout=types.NFHWC)
self.coin = ops.CoinFlip(probability=0.5)
self.brightness_val = ops.Uniform(
range=[1 - brightness, 1 + brightness])
self.contrast_val = ops.Uniform(range=[1 - contrast, 1 + contrast])
self.supervised = supervised
self.half = ops.Constant(fdata=0.5)
self.zero = ops.Constant(idata=0)
self.cast_to_long = ops.Cast(device='gpu', dtype=types.INT64)
if crop_size is not None:
H, W = crop_size
else:
# default
H, W = image_shape
# print('CONV MODE!!! {}'.format(conv_mode))
if conv_mode == '3d':
self.transpose = ops.Transpose(device="gpu", perm=[3, 0, 1, 2])
self.reshape = None
elif conv_mode == '2d':
self.transpose = ops.Transpose(device='gpu', perm=[0, 3, 1, 2])
self.reshape = ops.Reshape(device='gpu', shape=[-1, H, W])
self.validate = validate
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