def prepare_strong_aug(im,
dataset,
split,
mean,
sd,
eig_vals=None,
eig_vecs=None):
if "imagenet" in dataset:
# Train and test setups differ
train_size = cfg.TRAIN.IM_SIZE
if split == "train":
# Scale and aspect ratio then horizontal flip
im = transforms.random_sized_crop(im=im,
size=train_size,
area_frac=0.08)
im = transforms.horizontal_flip(im=im, p=0.5, order="HWC")
else:
# Scale and center crop
im = transforms.scale(cfg.TEST.IM_SIZE, im)
im = transforms.center_crop(train_size, im)
elif "cityscapes" in dataset:
train_size = cfg.TRAIN.IM_SIZE
if split == "train":
# Scale
random_scale = np.power(2, -1 + 2 * np.random.uniform())
random_size = int(max(train_size, cfg.TEST.IM_SIZE * random_scale))
im = transforms.scale(random_size, im)
# Crop
im = transforms.random_crop(im, train_size, order="HWC")
# Flip
im = transforms.horizontal_flip(im=im, p=0.5, order="HWC")
else:
# Scale
im = transforms.scale(cfg.TEST.IM_SIZE, im)
# Crop
im = transforms.center_crop(train_size, im)
im = transforms.HWC2CHW(im)
im = im / 255.0
if "cifar" not in dataset:
im = im[:, :, ::-1] # RGB -> BGR
# PCA jitter
if split == "train":
im = transforms.lighting(im, 0.1, eig_vals, eig_vecs)
# Color normalization
im = transforms.color_norm(im, mean, sd)
if "cifar" in dataset:
if split == "train":
if cfg.TASK == 'fix':
im = transforms.horizontal_flip(im=im, p=0.5)
im = transforms.random_crop(
im=im, size=cfg.TRAIN.IM_SIZE, pad_size=4
) # Best after color_norm because of zero padding
im = randaugment.random_augment(im)
else:
im = transforms.horizontal_flip(im=im, p=0.5)
im = transforms.random_crop(im=im,
size=cfg.TRAIN.IM_SIZE,
pad_size=4)
return im
def prepare_col(im, dataset, split, nbrs, mean, sd, eig_vals=None, eig_vecs=None):
if "cifar" not in dataset:
train_size = cfg.TRAIN.IM_SIZE
if split == "train":
if "imagenet" in dataset:
im = transforms.random_sized_crop(im=im, size=train_size, area_frac=0.08)
elif "cityscapes" in dataset:
random_scale = np.power(2, -1 + 2 * np.random.uniform())
random_size = int(max(train_size, cfg.TEST.IM_SIZE * random_scale))
im = transforms.scale(random_size, im)
im = transforms.random_crop(im, train_size, order="HWC")
else:
im = transforms.scale(cfg.TEST.IM_SIZE, im)
im = transforms.center_crop(train_size, im)
if split == "train":
im = transforms.horizontal_flip(im=im, p=0.5, order="HWC") # Best before rgb2lab because otherwise need to flip together with label
im_lab = color.rgb2lab(im.astype(np.uint8, copy=False))
im = transforms.HWC2CHW(im_lab[:, :, 0:1]).astype(np.float32, copy=False)
# Ad hoc normalization of the L channel
im = im / 100.0
im = im - np.mean(mean)
im = im / np.mean(sd)
label = nbrs.kneighbors(im_lab[:, :, 1:].reshape(-1, 2),
return_distance=False).reshape(im_lab.shape[0], im_lab.shape[1])
return im, label
def _prepare_im(self, im):
"""Prepares the image for network input."""
im = transforms.color_norm(im, _MEAN, _SD)
if self._split == "train":
im = transforms.horizontal_flip(im=im, p=0.5)
im = transforms.random_crop(im=im, size=cfg.TRAIN.IM_SIZE, pad_size=4)
return im
def _transform_image(self, image):
"""Transforms the image for network input."""
image = transforms.color_normalization(image, _MEAN, _SD)
if self._split == 'train':
image = transforms.horizontal_flip(image=image, prob=0.5)
image = transforms.random_crop(image=image, size=32, pad_size=4)
return image
def prepare_jig(im,
dataset,
split,
perms,
mean,
sd,
eig_vals=None,
eig_vecs=None):
if "cifar" not in dataset:
train_size = cfg.TRAIN.IM_SIZE
if split == "train":
if "imagenet" in dataset:
target_size = cfg.TEST.IM_SIZE
elif "cityscapes" in dataset:
random_scale = np.power(2, -1 + 2 * np.random.uniform())
target_size = int(
max(train_size, cfg.TEST.IM_SIZE * random_scale))
im = transforms.scale(target_size, im)
im = transforms.random_crop(im, train_size, order="HWC")
else:
im = transforms.scale(cfg.TEST.IM_SIZE, im)
im = transforms.center_crop(train_size, im)
if split == "train":
im = transforms.horizontal_flip(im=im, p=0.5, order="HWC")
if np.random.uniform() < cfg.TRAIN.GRAY_PERCENTAGE:
im = color.rgb2gray(im.astype(np.uint8, copy=False)) * 255.0
im = np.expand_dims(im, axis=2)
im = np.tile(im, (1, 1, 3))
im = transforms.HWC2CHW(im)
im = im / 255.0
if "cifar" not in dataset:
im = im[:, :, ::-1] # RGB -> BGR
# PCA jitter
if split == "train":
im = transforms.lighting(im, 0.1, eig_vals, eig_vecs)
# Color normalization
im = transforms.color_norm(im, mean, sd)
# Random permute
label = np.random.randint(len(perms))
perm = perms[label]
# Crop tiles
psz = int(cfg.TRAIN.IM_SIZE / cfg.JIGSAW_GRID) # Patch size
tsz = int(psz * 0.76) # Tile size; int(85 * 0.76) = 64
tiles = np.zeros((cfg.JIGSAW_GRID**2, 3, tsz, tsz)).astype(np.float32)
for i in range(cfg.JIGSAW_GRID):
for j in range(cfg.JIGSAW_GRID):
patch = im[:, psz * i:psz * (i + 1), psz * j:psz * (j + 1)]
# Gap
h = np.random.randint(psz - tsz + 1)
w = np.random.randint(psz - tsz + 1)
tile = patch[:, h:h + tsz, w:w + tsz]
# Normalization
mu, sigma = np.mean(tile), np.std(tile)
tile = tile - mu
tile = tile / sigma
tiles[perm[cfg.JIGSAW_GRID * i + j]] = tile
return tiles, label
def _prepare_im(self, im, transform=True):
# Train and test setups differ
train_size, test_size = cfg.TRAIN.IM_SIZE, cfg.TEST.IM_SIZE
if self._split == "train" and transform:
# For training use random_sized_crop, horizontal_flip, augment, lighting
im = transforms.random_sized_crop(im, train_size)
im = transforms.horizontal_flip(im, 0.5)
im = transforms.augment(im, cfg.TRAIN.AUGMENT)
im = transforms.lighting(im, cfg.TRAIN.PCA_STD, _EIG_VALS, _EIG_VECS)
else:
# For testing use scale and center crop
im = transforms.scale_and_center_crop(im, test_size, train_size)
if transform:
# For training and testing use color normalization
im = transforms.color_norm(im, _MEAN, _STD)
# Convert HWC/RGB/float to CHW/BGR/float format
im = np.ascontiguousarray(im[:, :, ::-1].transpose([2, 0, 1]))
return im
def _prepare_im(self, im):
"""Prepares the image for network input (HWC/BGR/int -> CHW/BGR/float)."""
# Convert HWC/BGR/int to HWC/RGB/float format for applying transforms
im = im[:, :, ::-1].astype(np.float32) / 255
# Train and test setups differ
train_size, test_size = cfg.TRAIN.IM_SIZE, cfg.TEST.IM_SIZE
if self._split == "train":
# For training use random_sized_crop, horizontal_flip, augment, lighting
im = transforms.random_sized_crop(im, train_size)
im = transforms.horizontal_flip(im, prob=0.5)
im = transforms.augment(im, cfg.TRAIN.AUGMENT)
im = transforms.lighting(im, cfg.TRAIN.PCA_STD, _EIG_VALS, _EIG_VECS)
else:
# For testing use scale and center crop
im = transforms.scale_and_center_crop(im, test_size, train_size)
# For training and testing use color normalization
im = transforms.color_norm(im, _MEAN, _STD)
# Convert HWC/RGB/float to CHW/BGR/float format
im = np.ascontiguousarray(im[:, :, ::-1].transpose([2, 0, 1]))
return im
def _prepare_im(self, im):
"""Prepares the image for network input."""
# Train and test setups differ
train_size = cfg.TRAIN.IM_SIZE
if self._split == "train":
# Scale and aspect ratio then horizontal flip
im = transforms.random_sized_crop(im=im, size=train_size, area_frac=0.08)
im = transforms.horizontal_flip(im=im, p=0.5, order="HWC")
else:
# Scale and center crop
im = transforms.scale(cfg.TEST.IM_SIZE, im)
im = transforms.center_crop(train_size, im)
# HWC -> CHW
im = im.transpose([2, 0, 1])
# [0, 255] -> [0, 1]
im = im / 255.0
# PCA jitter
if self._split == "train":
im = transforms.lighting(im, 0.1, _EIG_VALS, _EIG_VECS)
# Color normalization
im = transforms.color_norm(im, _MEAN, _SD)
return im
def _prepare_im(self, im):
"""Prepares the image for network input."""
# Train and test setups differ
if self._split == 'train':
# Scale and aspect ratio
im = transforms.random_sized_crop(image=im,
size=224,
area_frac=0.08)
# Horizontal flip
im = transforms.horizontal_flip(image=im, prob=0.5, order='HWC')
else:
# Scale and center crop
im = transforms.scale(256, im)
im = transforms.center_crop(224, im)
# HWC -> CHW
im = transforms.HWC2CHW(im)
# [0, 255] -> [0, 1]
im = im / 255.0
# PCA jitter
if self._split == 'train':
im = transforms.lighting(im, 0.1, _EIG_VALS, _EIG_VECS)
# Color normalization
im = transforms.color_normalization(im, _MEAN, _SD)
return im
