def __init__(self,
batch_size,
num_threads,
device_id,
data_dir,
crop,
size=256,
dali_cpu=False,
local_rank=0,
world_size=1):
super(HybridTrainPipe, self).__init__(batch_size,
num_threads,
device_id,
seed=12 + device_id)
dali_device = "gpu"
self.input = ops.FileReader(file_root=data_dir,
shard_id=local_rank,
num_shards=world_size,
random_shuffle=True)
self.decode = ops.ImageDecoder(device="mixed", output_type=types.RGB)
# self.res = ops.Resize(device="gpu", resize_x=size, resize_y=size, interp_type=types.INTERP_LINEAR)
self.res = ops.Resize(device="gpu",
resize_shorter=size,
interp_type=types.INTERP_LINEAR)
self.rescrop = ops.RandomResizedCrop(device="gpu",
size=crop,
random_area=[0.08, 1.25])
self.bc = ops.BrightnessContrast(device="gpu",
brightness=0.5,
contrast=0.6)
# Will flip vertically with prob of 0.1
self.vert_flip = ops.Flip(device='gpu', horizontal=0)
self.vert_coin = ops.CoinFlip(probability=0.4)
self.transform_source = ops.ExternalSource()
self.warp_keep_size = ops.WarpAffine(
device="gpu",
# size # keep original canvas size
interp_type=types.INTERP_LINEAR # use linear interpolation
)
# My workaround for Dali not supporting random affine transforms:
# a "synthetic random" warp affine transform.
# Rotate within a narrow range with probability of 0.075
self.rotate = ops.Rotate(device='gpu')
self.rotate_range = ops.Uniform(range=(-20.0, 20.0))
self.rotate_coin = ops.CoinFlip(probability=0.075)
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])
self.coin = ops.CoinFlip(probability=0.5)
print('DALI "{0}" variant'.format(dali_device))
def __init__(self, target_size, preproc_param, training=False):
self.training = training
mean = preproc_param.MEAN
std = preproc_param.STD
bri_delta = preproc_param.BRI_DELTA
hue_delta = preproc_param.HUE_DELTA
max_expand_ratio = preproc_param.MAX_EXPAND_RATIO
contrast_range = preproc_param.CONTRAST_RANGE
saturation_range = preproc_param.SATURATION_RANGE
crop_aspect_ratio = preproc_param.CROP_ASPECT_RATIO
crop_scale = preproc_param.CROP_SCALE
crop_attempts = preproc_param.CROP_ATTEMPTS
# decoder
self.decode_train = ops.ImageDecoderSlice(device="mixed", output_type=types.RGB)
self.decode_infer = ops.ImageDecoder(device="mixed", output_type=types.RGB)
# ssd crop
self.bbox_crop = ops.RandomBBoxCrop(
device="cpu",
bbox_layout="xyXY",
scaling=crop_scale,
aspect_ratio=crop_aspect_ratio,
allow_no_crop=True,
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
num_attempts=crop_attempts,
)
# color twist
self.uniform_con = ops.Uniform(range=contrast_range)
self.uniform_bri = ops.Uniform(
range=(1.0 - bri_delta / 256.0, 1.0 + bri_delta / 256.0)
)
self.uniform_sat = ops.Uniform(range=saturation_range)
self.uniform_hue = ops.Uniform(range=(-hue_delta, hue_delta))
self.hsv = ops.Hsv(device="gpu")
self.contrast = ops.BrightnessContrast(device="gpu")
# hflip
self.bbox_flip = ops.BbFlip(device="cpu", ltrb=True)
self.img_flip = ops.Flip(device="gpu")
self.coin_flip = ops.CoinFlip(probability=0.5)
# past
self.paste_pos = ops.Uniform(range=(0, 1))
self.paste_ratio = ops.Uniform(range=(1, max_expand_ratio))
self.paste = ops.Paste(device="gpu", fill_value=mean)
self.bbox_paste = ops.BBoxPaste(device="cpu", ltrb=True)
# resize and normalize
self.resize = ops.Resize(
device="gpu",
interp_type=types.DALIInterpType.INTERP_CUBIC,
resize_x=target_size[0],
resize_y=target_size[1],
save_attrs=True,
)
self.normalize = ops.CropMirrorNormalize(device="gpu", mean=mean, std=std)
def __init__(self,
file_root,
annotations_file,
batch_size=1,
device_id=0,
num_threads=4,
local_rank=0,
world_size=1):
super(HybridTrainPipe, self).__init__(batch_size,
num_threads,
device_id,
seed=42 + device_id)
self.reader = ops.COCOReader(
file_root=file_root,
annotations_file=annotations_file,
skip_empty=True,
shard_id=local_rank,
num_shards=world_size,
ratio=True,
ltrb=True,
shuffle_after_epoch=True,
pad_last_batch=True)
self.crop = ops.RandomBBoxCrop(
device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0],
ltrb=True,
allow_no_crop=True,
num_attempts=50)
self.bbflip = ops.BbFlip(device="cpu", ltrb=True)
self.roi_decode = ops.ImageDecoderSlice(device="mixed")
self.resize = ops.Resize(
device="gpu",
resize_x=300,
resize_y=300,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
self.hsv = ops.Hsv(device="gpu", dtype=types.FLOAT) # use float to avoid clipping and
# quantizing the intermediate result
self.bc = ops.BrightnessContrast(device="gpu",
contrast_center=128, # input is in float, but in 0..255 range
dtype=types.UINT8)
self.cmnp = ops.CropMirrorNormalize(
device="gpu",
mean=[104., 117., 123.],
std=[1., 1., 1.],
dtype=types.FLOAT,
output_layout=types.NCHW,
pad_output=False)
self.rng1 = ops.Uniform(range=[0.5, 1.5])
self.rng2 = ops.Uniform(range=[0.875, 1.125])
self.rng3 = ops.Uniform(range=[-0.5, 0.5])
self.coin = ops.CoinFlip(probability=0.5)
self.build()
def __new__(cls, **kwargs):
"""Create a ``BrightnessContrast`` operator.
Returns
-------
nvidia.dali.ops.BrightnessContrast
The operator.
"""
return ops.BrightnessContrast(device=context.get_device_type(), **kwargs)
def __init__(self, batch_size, num_threads, path, training, annotations, world, device_id, mean, std, resize,
max_size, stride, rotate_augment=False,
augment_brightness=0.0,
augment_contrast=0.0, augment_hue=0.0,
augment_saturation=0.0):
super().__init__(batch_size=batch_size, num_threads=num_threads, device_id=device_id,
prefetch_queue_depth=num_threads, seed=42)
self.path = path
self.training = training
self.stride = stride
self.iter = 0
self.rotate_augment = rotate_augment
self.augment_brightness = augment_brightness
self.augment_contrast = augment_contrast
self.augment_hue = augment_hue
self.augment_saturation = augment_saturation
self.reader = ops.COCOReader(annotations_file=annotations, file_root=path, num_shards=world,
shard_id=torch.cuda.current_device(),
ltrb=True, ratio=True, shuffle_after_epoch=True, save_img_ids=True)
self.decode_train = ops.ImageDecoderSlice(device="mixed", output_type=types.RGB)
self.decode_infer = ops.ImageDecoder(device="mixed", output_type=types.RGB)
self.bbox_crop = ops.RandomBBoxCrop(device='cpu', ltrb=True, scaling=[0.3, 1.0],
thresholds=[0.1, 0.3, 0.5, 0.7, 0.9])
self.bbox_flip = ops.BbFlip(device='cpu', ltrb=True)
self.img_flip = ops.Flip(device='gpu')
self.coin_flip = ops.CoinFlip(probability=0.5)
self.bc = ops.BrightnessContrast(device='gpu')
self.hsv = ops.Hsv(device='gpu')
# Random number generation for augmentation
self.brightness_dist = ops.NormalDistribution(mean=1.0, stddev=augment_brightness)
self.contrast_dist = ops.NormalDistribution(mean=1.0, stddev=augment_contrast)
self.hue_dist = ops.NormalDistribution(mean=0.0, stddev=augment_hue)
self.saturation_dist = ops.NormalDistribution(mean=1.0, stddev=augment_saturation)
if rotate_augment:
raise RuntimeWarning("--augment-rotate current has no effect when using the DALI data loader.")
if isinstance(resize, list): resize = max(resize)
self.rand_resize = ops.Uniform(range=[resize, float(max_size)])
self.resize_train = ops.Resize(device='gpu', interp_type=types.DALIInterpType.INTERP_CUBIC, save_attrs=True)
self.resize_infer = ops.Resize(device='gpu', interp_type=types.DALIInterpType.INTERP_CUBIC,
resize_longer=max_size, save_attrs=True)
padded_size = max_size + ((self.stride - max_size % self.stride) % self.stride)
self.pad = ops.Paste(device='gpu', fill_value=0, ratio=1.1, min_canvas_size=padded_size, paste_x=0, paste_y=0)
self.normalize = ops.CropMirrorNormalize(device='gpu', mean=mean, std=std, crop=(padded_size, padded_size),
crop_pos_x=0, crop_pos_y=0)
def __init__(self,
p: float = .5,
brightness_limit: Union[List, float] = 0.5,
contrast_limit: Union[List, float] = 0.5):
"""Initialization
Args:
p (float, optional): Probability to apply this transformation. Defaults to .5.
brightness_limit (Union[List,float], optional): Factor multiplier range for changing brightness in [min,max] value format. If provided as a single float, the range will be 1 + (-limit, limit). Defaults to 0.5.
contrast_limit (Union[List,float], optional): Factor multiplier range for changing contrast in [min,max] value format. If provided as a single float, the range will be 1 + (-limit, limit). Defaults to 0.5.
"""
brightness_limit = _check_and_convert_limit_value(brightness_limit)
contrast_limit = _check_and_convert_limit_value(contrast_limit)
self.brightness_uniform = ops.Uniform(range=brightness_limit)
self.contrast_uniform = ops.Uniform(range=contrast_limit)
self.random_brightness_contrast = ops.BrightnessContrast(device='gpu')
self.rng = ops.CoinFlip(probability=p)
self.bool = ops.Cast(dtype=types.DALIDataType.BOOL)
def __init__(self, default_boxes, args, seed):
super(COCOPipeline, self).__init__(batch_size=args.batch_size,
device_id=args.local_rank,
num_threads=args.num_workers,
seed=seed)
try:
shard_id = torch.distributed.get_rank()
num_shards = torch.distributed.get_world_size()
except RuntimeError:
shard_id = 0
num_shards = 1
self.input = ops.COCOReader(file_root=args.train_coco_root,
annotations_file=args.train_annotate,
skip_empty=True,
shard_id=shard_id,
num_shards=num_shards,
ratio=True,
ltrb=True,
random_shuffle=False,
shuffle_after_epoch=True)
self.decode = ops.ImageDecoder(device="cpu", output_type=types.RGB)
# Augumentation techniques
self.crop = ops.RandomBBoxCrop(device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0],
ltrb=True,
allow_no_crop=True,
num_attempts=1)
self.slice = ops.Slice(device="cpu")
self.hsv = ops.Hsv(
device="gpu", dtype=types.FLOAT) # use float to avoid clipping and
# quantizing the intermediate result
self.bc = ops.BrightnessContrast(
device="gpu",
contrast_center=128, # input is in float, but in 0..255 range
dtype=types.UINT8)
self.resize = ops.Resize(
device="cpu",
resize_x=300,
resize_y=300,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
dtype = types.FLOAT16 if args.fp16 else types.FLOAT
self.normalize = ops.CropMirrorNormalize(
device="gpu",
crop=(300, 300),
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255],
mirror=0,
dtype=dtype,
output_layout=types.NCHW,
pad_output=False)
# Random variables
self.rng1 = ops.Uniform(range=[0.5, 1.5])
self.rng2 = ops.Uniform(range=[0.875, 1.125])
self.rng3 = ops.Uniform(range=[-0.5, 0.5])
self.flip = ops.Flip(device="cpu")
self.bbflip = ops.BbFlip(device="cpu", ltrb=True)
self.flip_coin = ops.CoinFlip(probability=0.5)
self.box_encoder = ops.BoxEncoder(device="cpu",
criteria=0.5,
anchors=default_boxes.as_ltrb_list())
def __init__(self, args, device_id, file_root, annotations_file):
super(DetectionPipeline,
self).__init__(batch_size=args.batch_size,
num_threads=args.num_workers,
device_id=device_id,
prefetch_queue_depth=args.prefetch,
seed=args.seed)
# Reading COCO dataset
self.input = ops.readers.COCO(file_root=file_root,
annotations_file=annotations_file,
shard_id=device_id,
num_shards=args.num_gpus,
ratio=True,
ltrb=True,
random_shuffle=True)
self.decode_cpu = ops.decoders.Image(device="cpu",
output_type=types.RGB)
self.decode_crop = ops.decoders.ImageSlice(device="cpu",
output_type=types.RGB)
self.decode_gpu = ops.decoders.Image(device="mixed",
output_type=types.RGB,
hw_decoder_load=0)
self.decode_gpu_crop = ops.decoders.ImageSlice(device="mixed",
output_type=types.RGB,
hw_decoder_load=0)
self.ssd_crop = ops.SSDRandomCrop(device="cpu",
num_attempts=1,
seed=args.seed)
self.random_bbox_crop = ops.RandomBBoxCrop(
device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0],
bbox_layout="xyXY",
seed=args.seed)
self.slice_cpu = ops.Slice(device="cpu")
self.slice_gpu = ops.Slice(device="gpu")
self.resize_cpu = ops.Resize(
device="cpu",
resize_x=300,
resize_y=300,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
self.resize_gpu = ops.Resize(
device="gpu",
resize_x=300,
resize_y=300,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
mean = [0.485 * 255, 0.456 * 255, 0.406 * 255]
std = [0.229 * 255, 0.224 * 255, 0.225 * 255]
crop_size = (300, 300)
self.normalize_cpu = ops.CropMirrorNormalize(device="cpu",
crop=crop_size,
mean=mean,
std=std,
mirror=0,
dtype=types.FLOAT)
self.normalize_gpu = ops.CropMirrorNormalize(device="gpu",
crop=crop_size,
mean=mean,
std=std,
mirror=0,
dtype=types.FLOAT)
self.twist_cpu = ops.ColorTwist(device="cpu")
self.twist_gpu = ops.ColorTwist(device="gpu")
self.hsv_cpu = ops.Hsv(device="cpu", dtype=types.FLOAT)
self.hsv_gpu = ops.Hsv(device="gpu", dtype=types.FLOAT)
self.bc_cpu = ops.BrightnessContrast(device="cpu",
dtype=types.UINT8,
contrast_center=128)
self.bc_gpu = ops.BrightnessContrast(device="gpu",
dtype=types.UINT8,
contrast_center=128)
self.flip_cpu = ops.Flip(device="cpu")
self.bbox_flip_cpu = ops.BbFlip(device="cpu", ltrb=True)
self.flip_gpu = ops.Flip(device="gpu")
self.bbox_flip_gpu = ops.BbFlip(device="gpu", ltrb=True)
default_boxes = coco_anchors()
self.box_encoder_cpu = ops.BoxEncoder(device="cpu",
criteria=0.5,
anchors=default_boxes)
self.box_encoder_gpu = ops.BoxEncoder(device="gpu",
criteria=0.5,
anchors=default_boxes)
self.box_encoder_cpu_offsets = ops.BoxEncoder(
device="cpu",
criteria=0.5,
offset=True,
scale=2,
stds=[0.1, 0.1, 0.2, 0.2],
anchors=default_boxes)
self.box_encoder_gpu_offsets = ops.BoxEncoder(
device="gpu",
criteria=0.5,
offset=True,
scale=2,
stds=[0.1, 0.1, 0.2, 0.2],
anchors=default_boxes)
# Random variables
self.saturation_rng = ops.random.Uniform(range=[0.8, 1.2])
self.contrast_rng = ops.random.Uniform(range=[0.5, 1.5])
self.brighness_rng = ops.random.Uniform(range=[0.875, 1.125])
self.hue_rng = ops.random.Uniform(range=[-45, 45])
def __init__(self, batch_size, num_threads, device_id, seed, image_dir):
super(BrightnessPipeline, self).__init__(batch_size, num_threads,
device_id, seed, image_dir)
self.brightness = ops.BrightnessContrast(device="gpu", brightness=0.5)
def __init__(self, train=False, batch_size=16, workers=4, size=512):
# TODO: support size as tuple
local_rank, world_size = env_rank(), env_world_size()
super().__init__(batch_size, workers, local_rank, seed=42)
split_str = "train" if train else "val"
self.input = ops.COCOReader(
file_root=f"{DATA_DIR}/{split_str}2017",
annotations_file=
f"{DATA_DIR}/annotations/instances_{split_str}2017.json",
shard_id=local_rank,
num_shards=world_size,
ratio=True, # want bbox in [0, 1]
ltrb=True, #
random_shuffle=train,
save_img_ids=True, # Need ids for evaluation
skip_empty=
True, # skips images without objects. not sure if we want to do so
)
self.bbox_crop = ops.RandomBBoxCrop(
device="cpu", # gpu is not supported (and not needed actually)
bbox_layout="xyXY", # same as 'ltrb'
scaling=[0.3, 1.0],
# adding 0.0 to thr instead of `allow_no_crop`
thresholds=[0.0, 0.1, 0.3, 0.5, 0.7, 0.9],
)
if train:
self.decode = ops.ImageDecoderSlice(device="mixed",
output_type=types.RGB)
else:
self.decode = ops.ImageDecoder(device="mixed",
output_type=types.RGB)
self.resize = ops.Resize(device="gpu",
interp_type=types.INTERP_CUBIC,
resize_longer=size,
save_attrs=True)
self.resize = ops.Resize(device="gpu",
interp_type=types.INTERP_CUBIC,
resize_longer=size)
self.bbox_flip = ops.BbFlip(device="cpu", ltrb=True)
self.img_flip = ops.Flip(device="gpu")
# color augmentations
self.bc = ops.BrightnessContrast(device="gpu")
self.hsv = ops.Hsv(device="gpu")
# pad to match output stride
self.pad = ops.Pad(device="gpu",
fill_value=0,
axes=(1, 2),
shape=(size, size))
self.normalize = ops.CropMirrorNormalize(
device="gpu",
# Imagenet mean and std
mean=[0.485 * 255, 0.456 * 255, 0.406 * 255],
std=[0.229 * 255, 0.224 * 255, 0.225 * 255],
# mean=[0, 0, 0],
# std=[1, 1, 1],
output_dtype=types.FLOAT,
output_layout=types.NCHW,
)
# TODO: add Jitter aug
# Random number generation for augmentation
self.coin_flip = ops.CoinFlip(probability=0.5)
self.rng1 = ops.Uniform(range=[0, 1])
self.rng2 = ops.Uniform(range=[0.85, 1.15])
self.rng3 = ops.Uniform(range=[-15, 15])
self.train = train
def __init__(
self,
train=False,
batch_size=16,
size=384,
num_threads=4,
device_id=0
):
super(ExternalSourcePipeline, self).__init__(
batch_size, num_threads, device_id, seed=42
)
self.eii = iter(
ExternalInputIterator(train, batch_size)
)
self.images_input = ops.ExternalSource()
self.masks_input = ops.ExternalSource()
if train:
fixed_area = (size / 784)**2
self.decode = ops.ImageDecoderRandomCrop(
device="mixed",
random_area=[fixed_area*0.7, fixed_area*1.3],
random_aspect_ratio=[0.7, 1.3],
)
else:
self.decode = ops.ImageDecoderCrop(
device="mixed",
crop=(size, size)
)
self.resize = ops.Resize(
device="gpu",
interp_type=types.INTERP_TRIANGULAR,
resize_x=size,
resize_y=size
)
self.mask_resize = ops.Resize(
device="gpu",
interp_type=types.INTERP_NN,
resize_x=size,
resize_y=size
)
self.normalize = ops.CropMirrorNormalize(
device="gpu",
mean=[0.5 * 255], # 0.456 * 255, 0.406 * 255],
std=[0.5 * 255], # , 0.224 * 255, 0.225 * 255],
output_layout=types.NCHW
)
self.mask_normalize = ops.CropMirrorNormalize(
device="gpu",
mean=[0],
std=[255],
output_layout=types.NCHW,
image_type=types.GRAY,
)
# extra augmentations
self.to_gray = ops.ColorSpaceConversion(
device="gpu", image_type=types.RGB, output_type=types.GRAY
)
self.contrast = ops.BrightnessContrast(device="gpu")
self.hsv = ops.Hsv(device="gpu")
self.jitter = ops.Jitter(device ="gpu")
# self.rng1 = ops.Uniform(range=[0, 1])
self.rng2 = ops.Uniform(range=[0.8,1.2])
self.rng3 = ops.Uniform(range=[-30, 30]) # for hue
self.coin03 = ops.CoinFlip(probability=0.3)
self.train = train
