def __call__(self,
img: Image.Image,
boxes: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None):
if random() < self._prob:
brightness_factor = uniform(self._min, self._max_brightness_factor)
img = F.adjust_brightness(img, brightness_factor)
contrast_factor = uniform(self._min, self._max_contrast_factor)
hue_factor = uniform(-self._min, self._max_hue_factor)
saturation_factor = uniform(self._min, self._max_saturation_factor)
# 50% chance applying contrast followed by hue and saturation or vice versa
if random() < 0.5:
if random() < self._prob:
img = F.adjust_contrast(img, contrast_factor)
if random() < self._prob:
img = img.convert("HSV")
img = F.adjust_hue(img, hue_factor)
img = img.convert("RGB")
img = F.adjust_saturation(img, saturation_factor)
else:
if random() < self._prob:
img = img.convert("HSV")
img = F.adjust_hue(img, hue_factor)
img = img.convert("RGB")
img = F.adjust_saturation(img, saturation_factor)
if random() < self._prob:
img = F.adjust_contrast(img, contrast_factor)
return img, boxes, labels
def __getitem__(self, index):
if self.mode == 1:
# mode =1 为预测时使用，会从所有WSI文件中返回全部的region的图像
slideIDX = self.slideIDX[index]
coord = self.grid[index]
img = self.slides[slideIDX].read_region(coord,self.level,(self.patch_size[slideIDX],\
self.patch_size[slideIDX])).convert('RGB')
if img.size != (224, 224):
img = img.resize((224, 224), Image.BILINEAR)
if self.transform is not None:
img = self.transform(img)
return img
elif self.mode == 2:
# mode =2 为训练时使用，只会根据指定的index(经过上一轮MIL过程得出) \
# 从全部WSI文件中筛选对应的坐标列表,返回相应的训练图像和label
slideIDX, coord, target, h_value = self.t_data[index]
img = self.slides[slideIDX].read_region(coord,self.level,(self.patch_size[slideIDX],\
self.patch_size[slideIDX])).convert('RGB')
if h_value > 0:
hue_factor = random.uniform(h_value, 0.1)
elif h_value == 0:
hue_factor = random.uniform(0, 0)
elif h_value < 0:
hue_factor = random.uniform(-0.1, h_value)
img = functional.adjust_hue(img, hue_factor)
# 只有在训练模式下才进行H通道变换的颜色增强方法
# 如果在maketraindata方法设置采样复制,那么就会针对h_value的值进行不同方向的hue_factor生成,\
# 从而达到复制的样本和原来的样本有不一样的增强的效果
if img.size != (224, 224):
img = img.resize((224, 224), Image.BILINEAR)
if self.transform is not None:
img = self.transform(img)
return img, target
elif self.mode == 3 and self.top_k is not None and self.feature_extract_model is not None:
k_value = int(len(self.top_k) / len(self.targets))
#通过上述方法得出实际的top k数值,省下不必要的外部传参
h_trans = np.random.binomial(1, 0.5)
for j in range(k_value):
coord = self.grid[self.top_k[index * k_value + j]]
img = self.slides[index].read_region(coord,self.level,(self.patch_size[index],\
self.patch_size[index])).convert('RGB')
if img.size != (224, 224):
img = img.resize((224, 224), Image.BILINEAR)
if h_trans:
#每一张slide由随机过程决定下属的gird的k张截图是否进行颜色变换,\
# 每张slide的截图要么全部不变换要么全部变换,防止出现同一个slide下有些截图变换有些截图不变换
hue_factor = random.uniform(-0.05, 0.05)
img = functional.adjust_hue(img, hue_factor)
img = self.transform(img).unsqueeze(0)
if j == 0:
feature = self.feature_extract_model(img.cuda())
#单个img生成的feature的shape是torch.Size([1,512]),512是resnet34全连接层前的特征维度
else:
feature = torch.cat(
(feature, self.feature_extract_model(img.cuda())), 0)
#在上述循环中,会将k个feature纵向叠加在一起,变成torch.Size([k,512])
return feature.view(-1, feature.shape[1]), self.targets[index]
def __call__(self, sample):
if np.random.random() < 0.5:
hue = np.random.uniform(-0.1, 0.1)
sample['left'] = F.adjust_hue(sample['left'], hue)
sample['right'] = F.adjust_hue(sample['right'], hue)
return sample
def CustomTransform(LR, HR, adjust_brightness, adjust_contrast,
adjust_saturation, adjust_hue, adjust_gamma, rotate, hflip,
vflip):
"""
to execute data augmentation.
except :param LR and :param HR, all other params accept bool args.
"""
if adjust_brightness:
brightness_factor = random.gauss(1, 0.05)
while brightness_factor > 2 or brightness_factor < 0:
brightness_factor = random.gauss(1, 0.05)
HR = f.adjust_brightness(HR, brightness_factor)
LR = f.adjust_brightness(LR, brightness_factor)
if adjust_contrast:
contrast_factor = random.gauss(1, 0.05)
while contrast_factor > 2 or contrast_factor < 0:
contrast_factor = random.gauss(1, 0.05)
HR = f.adjust_contrast(HR, contrast_factor)
LR = f.adjust_contrast(LR, contrast_factor)
if adjust_saturation:
saturation_factor = random.gauss(1, 0.05)
while saturation_factor > 2 or saturation_factor < 0:
saturation_factor = random.gauss(1, 0.05)
HR = f.adjust_saturation(HR, saturation_factor)
LR = f.adjust_saturation(LR, saturation_factor)
if adjust_hue:
hue_factor = random.gauss(0, 0.025)
while hue_factor > 0.5 or hue_factor < -0.5:
hue_factor = random.gauss(0, 0.025)
HR = f.adjust_hue(HR, hue_factor)
LR = f.adjust_hue(LR, hue_factor)
if adjust_gamma:
gamma = random.gauss(1, 0.025)
while gamma > 0.5 or gamma < -0.5:
gamma = random.gauss(0, 0.025)
HR = f.adjust_gamma(HR, gamma)
LR = f.adjust_gamma(LR, gamma)
if rotate:
angle = random.choice([-90, 0, 90])
HR = f.rotate(HR, angle)
LR = f.rotate(LR, angle)
if hflip:
flip = random.choice([True, False])
if flip:
HR = f.hflip(HR)
LR = f.hflip(LR)
if vflip:
flip = random.choice([True, False])
if flip:
HR = f.vflip(HR)
LR = f.vflip(LR)
return LR, HR
def __call__(self, seq_blur, seq_clear):
seq_blur = [Image.fromarray(np.uint8(img)) for img in seq_blur]
seq_clear = [Image.fromarray(np.uint8(img)) for img in seq_clear]
if self.brightness > 0:
brightness_factor = np.random.uniform(max(0, 1 - self.brightness), 1 + self.brightness)
seq_blur = [F.adjust_brightness(img, brightness_factor) for img in seq_blur]
seq_clear = [F.adjust_brightness(img, brightness_factor) for img in seq_clear]
if self.contrast > 0:
contrast_factor = np.random.uniform(max(0, 1 - self.contrast), 1 + self.contrast)
seq_blur = [F.adjust_contrast(img, contrast_factor) for img in seq_blur]
seq_clear = [F.adjust_contrast(img, contrast_factor) for img in seq_clear]
if self.saturation > 0:
saturation_factor = np.random.uniform(max(0, 1 - self.saturation), 1 + self.saturation)
seq_blur = [F.adjust_saturation(img, saturation_factor) for img in seq_blur]
seq_clear = [F.adjust_saturation(img, saturation_factor) for img in seq_clear]
if self.hue > 0:
hue_factor = np.random.uniform(-self.hue, self.hue)
seq_blur = [F.adjust_hue(img, hue_factor) for img in seq_blur]
seq_clear = [F.adjust_hue(img, hue_factor) for img in seq_clear]
seq_blur = [np.asarray(img) for img in seq_blur]
seq_clear = [np.asarray(img) for img in seq_clear]
seq_blur = [img.clip(0,255) for img in seq_blur]
seq_clear = [img.clip(0,255) for img in seq_clear]
return seq_blur, seq_clear
def custom_transform(self, image, segmentation):
# 0.5 probability of performing one transformation
if random.random() > 0.5:
p = random.random()
# Rotate
if p < 0.3:
angle = random.randint(-30, 30)
image = TF.rotate(image, angle)
segmentation = TF.rotate(segmentation, angle)
# Horizontal flip
elif p < 0.6:
TF.hflip(image)
TF.hflip(segmentation)
# Vertical flip
elif p < 0.9:
TF.vflip(image)
TF.vflip(segmentation)
# Colours
else:
TF.adjust_brightness(image, 1.4)
TF.adjust_hue(image, 0.3)
TF.adjust_contrast(image, 1.3)
TF.adjust_saturation(image, 0.7)
return image, segmentation
def data_augmentation(x, y, mode='train'):
"""
x, y: np array, [B, C, H, W]
"""
x_list = list()
y_list = list()
for i in range(x.shape[0]):
# gamma correction
xi = skimage.exposure.adjust_gamma(x[i, ...], 0.4)
yi = skimage.exposure.adjust_gamma(y[i, ...], 0.4)
# color transfer
if mode == 'train':
xi = color_transform(xi, T)
yi = color_transform(yi, T)
xi = Image.fromarray(xi)
yi = Image.fromarray(yi)
if mode == 'train':
pick = random.randint(0, 4)
if pick == 0:
# random brightness
brightness_factor = 1.0 + random.uniform(0, 0.3)
xi = ttf.adjust_brightness(xi, brightness_factor)
yi = ttf.adjust_brightness(yi, brightness_factor)
elif pick == 1:
# random saturation
saturation_factor = 1.0 + random.uniform(-0.2, 0.5)
xi = ttf.adjust_saturation(xi, saturation_factor)
yi = ttf.adjust_saturation(yi, saturation_factor)
elif pick == 2:
# random hue
hue_factor = random.uniform(-0.2, 0.2)
xi = ttf.adjust_hue(xi, hue_factor)
yi = ttf.adjust_hue(yi, hue_factor)
elif pick == 3:
# random contrast
contrast_factor = 1.0 + random.uniform(-0.2, 0.4)
xi = ttf.adjust_contrast(xi, contrast_factor)
yi = ttf.adjust_contrast(yi, contrast_factor)
elif pick == 4:
# random swap color channel
permute = np.random.permutation(3)
xi = np.array(xi)
yi = np.array(yi)
xi = xi[..., permute]
yi = yi[..., permute]
xi = np.clip(np.array(xi) / 255.0, 0, 1.0)
yi = np.clip(np.array(yi) / 255.0, 0, 1.0)
x_list.append(xi)
y_list.append(yi)
x_ret = torch.tensor(np.stack(x_list, axis=0), dtype=torch.float)
y_ret = torch.tensor(np.stack(y_list, axis=0), dtype=torch.float)
x_ret = normalize_lf(x_ret)
y_ret = normalize_lf(y_ret)
return x_ret.to(device), y_ret.to(device)
def __call__(self, image, ground_truth):
if self.train:
# shape augmentations
# left-right mirroring
if np.random.random() > 0.5:
image = TF.hflip(image)
ground_truth = TF.hflip(ground_truth)
# up-down mirroring
if np.random.random() > 0.5:
image = TF.vflip(image)
ground_truth = TF.vflip(ground_truth)
# random rotation
angle = np.random.uniform(-180, 180)
image = TF.rotate(image, angle, expand=True)
ground_truth = TF.rotate(ground_truth, angle, expand=True)
# center crop
center_crop = transforms.CenterCrop(int(self.output_size * 1.5))
image = center_crop(image)
ground_truth = center_crop(ground_truth)
# random resized crop
if np.random.random() > 0.2:
# random crop
i, j, h, w = transforms.RandomResizedCrop.get_params(image, scale=(0.2, 0.9), ratio=(1, 1))
image = TF.resized_crop(image, i, j, h, w, size=int(self.output_size))
ground_truth = TF.resized_crop(ground_truth, i, j, h, w, size=int(self.output_size))
# random crop without resize
else:
i, j, h, w = transforms.RandomCrop.get_params(image, output_size=(self.output_size, self.output_size))
image = TF.crop(image, i, j, h, w)
ground_truth = TF.crop(ground_truth, i, j, h, w)
# color augmentations
for col_aug in [TF.adjust_contrast, TF.adjust_brightness, TF.adjust_saturation, TF.adjust_gamma]:
if np.random.random() > 0.5:
adjust_factor = np.random.uniform(0.5, 1.5)
image = col_aug(image, adjust_factor)
ground_truth = col_aug(ground_truth, adjust_factor)
if np.random.random() > 0.5:
hue_factor = np.random.uniform(-0.15, 0.15)
image = TF.adjust_hue(image, hue_factor)
ground_truth = TF.adjust_hue(ground_truth, hue_factor)
else:
center_crop = transforms.CenterCrop(self.output_size)
image = center_crop(image)
ground_truth = center_crop(ground_truth)
# change locations to tensor
ground_truth = TF.normalize(TF.to_tensor(ground_truth), [0.5] * 3, [0.25] * 3)
image = TF.normalize(TF.to_tensor(image), [0.5] * 4, [0.25] * 4)
return image, ground_truth
def selective_color_distort(self, image, mask):
# make skin mask using face, neck, and ear segments
skin = mask[0] + mask[16] + mask[6] + mask[7]
# ensure there's no clipping
skin[skin > 1] = 1
skin = skin.repeat((3, 1, 1))
# torch transforms only work on PIL images, so first
# the image tensors must be converted to PIL images
background = TF.to_pil_image(image)
face = TF.to_pil_image(image)
# generate parameters for color jitter
# taken from https://arxiv.org/pdf/2002.05709.pdf page 12
jitter_params = torch.rand(4) * torch.tensor([.8, .8, .8, .2])
if torch.rand(1).item() > .2:
# apply transforms to background
background = TF.adjust_brightness(background,
.2 + jitter_params[0] * 2)
background = TF.adjust_contrast(background,
.2 + jitter_params[1] * 2)
background = TF.adjust_saturation(background,
.2 + jitter_params[2] * 2)
background = TF.adjust_hue(background, jitter_params[3] * 2 - .2)
# apply identical brightness/contrast/saturation transform to face
face = TF.adjust_brightness(face, .2 + jitter_params[0] * 2)
face = TF.adjust_saturation(face, .2 + jitter_params[2] * 2)
# only apply 50% of the contrast transform to the face
face = TF.adjust_contrast(face, .4 + jitter_params[1])
# only apply 25% of the hue transform to the face
face = TF.adjust_hue(face, jitter_params[3] * .5 - .05)
# note: the possibility of neither color jitter or grayscale or both
# color jitter and grayscale is intended behavior. the random
# draws are meant to be different for these two if statements
if torch.rand(1).item() > .8:
background = TF.to_grayscale(background, num_output_channels=3)
face = TF.to_grayscale(face, num_output_channels=3)
# convert PIL images back to tensors
background = TF.to_tensor(background)
face = TF.to_tensor(face)
# construct final image as convex combination of face pixels
# and mask pixels. where the mask has higher confidence (closer
# to 1), the face image will be used
distorted = (1 - skin) * background + skin * face
return distorted.type(torch.float32)
def __call__(self, input, target):
#Horizontal flip
if h_flip:
flip = random.random() < 0.5
if flip:
input=input.transpose(Image.FLIP_LEFT_RIGHT)
target = target.transpose(Image.FLIP_LEFT_RIGHT)
#Vertical flip
if v_flip:
flip = random.random() < 0.5
if flip:
input = input.transpose(Image.FLIP_TOP_BOTTOM)
target = target.transpose(Image.FLIP_TOP_BOTTOM)
#RandomCrop
if Random_crop:
i, j, h, w = transforms.RandomCrop.get_params(input, output_size=(224, 224))
input = TF.crop(input, i, j, h, w)
target = TF.crop(target, i, j, h, w)
#RandomRotation
i = transforms.RandomRotation.get_params(angle)
input = TF.rotate(input, i)
target = TF.rotate(target, i)
#ColorJitter
b,c,s,h=transforms.ColorJitter.get_params(ColorJitter[0],ColorJitter[1],ColorJitter[2],ColorJitter[3])
input=TF.adjust_brightness(input,b)
input=TF.adjust_contrast(input,c)
input = TF.adjust_saturation(input, s)
input = TF.adjust_hue(input, h)
target=TF.adjust_brightness(target,b)
target=TF.adjust_contrast(target,c)
target = TF.adjust_saturation(target, s)
target = TF.adjust_hue(target, h)
return input, target
def forward(self, img, mask):
"""
Args:
img (PIL Image or Tensor): Input image.
Returns:
PIL Image or Tensor: Color jittered image.
"""
fn_idx = torch.randperm(4)
for fn_id in fn_idx:
if fn_id == 0 and self.brightness is not None:
brightness = self.brightness
brightness_factor = torch.tensor(1.0).uniform_(
brightness[0], brightness[1]).item()
img = F.adjust_brightness(img, brightness_factor)
if fn_id == 1 and self.contrast is not None:
contrast = self.contrast
contrast_factor = torch.tensor(1.0).uniform_(
contrast[0], contrast[1]).item()
img = F.adjust_contrast(img, contrast_factor)
if fn_id == 2 and self.saturation is not None:
saturation = self.saturation
saturation_factor = torch.tensor(1.0).uniform_(
saturation[0], saturation[1]).item()
img = F.adjust_saturation(img, saturation_factor)
if fn_id == 3 and self.hue is not None:
hue = self.hue
hue_factor = torch.tensor(1.0).uniform_(hue[0], hue[1]).item()
img = F.adjust_hue(img, hue_factor)
return img, mask
def __call__(self, inputs):
inputs = [Image.fromarray(np.uint8(inp)) for inp in inputs]
if self.brightness > 0:
brightness_factor = np.random.uniform(max(0, 1 - self.brightness),
1 + self.brightness)
inputs = [
F.adjust_brightness(inp, brightness_factor) for inp in inputs
]
if self.contrast > 0:
contrast_factor = np.random.uniform(max(0, 1 - self.contrast),
1 + self.contrast)
inputs = [
F.adjust_contrast(inp, contrast_factor) for inp in inputs
]
if self.saturation > 0:
saturation_factor = np.random.uniform(max(0, 1 - self.saturation),
1 + self.saturation)
inputs = [
F.adjust_saturation(inp, saturation_factor) for inp in inputs
]
if self.hue > 0:
hue_factor = np.random.uniform(-self.hue, self.hue)
inputs = [F.adjust_hue(inp, hue_factor) for inp in inputs]
inputs = [np.asarray(inp) for inp in inputs]
inputs = [inp.clip(0, 255) for inp in inputs]
return inputs
def augment_PIL(self, img, labels):
if np.random.rand() > 0.4:
img = tvf.adjust_brightness(img, uniform(0.3, 1.5))
if np.random.rand() > 0.7:
factor = 2**uniform(-1, 1)
img = tvf.adjust_contrast(img, factor) # 0.5 ~ 2
if np.random.rand() > 0.7:
img = tvf.adjust_hue(img, uniform(-0.1, 0.1))
if np.random.rand() > 0.6:
factor = uniform(0, 2)
if factor > 1:
factor = 1 + uniform(0, 2)
img = tvf.adjust_saturation(img, factor) # 0 ~ 3
if np.random.rand() > 0.5:
img = tvf.adjust_gamma(img, uniform(0.5, 3))
# horizontal flip
if np.random.rand() > 0.5:
img, labels = augUtils.hflip(img, labels)
# vertical flip
if np.random.rand() > 0.5:
img, labels = augUtils.vflip(img, labels)
# # random rotation
rand_degree = np.random.rand() * 360
if self.coco:
img, labels = augUtils.rotate(img,
rand_degree,
labels,
expand=True)
else:
img, labels = augUtils.rotate(img,
rand_degree,
labels,
expand=False)
return img, labels
def transform(self, images):
if self.use_augmentation:
# Flip all
if np.random.rand() > 0.5:
images = [TF.hflip(i) for i in images]
# Change hue of RGB images
if np.random.rand() > 0.5:
hue_val = np.random.rand() - 0.5 # random val [-0.5, 0.5]
images = [
TF.adjust_hue(i, hue_val) if i.mode == "RGB" else i
for i in images
]
# Change saturation of RGB images
if np.random.rand() > 0.5:
sat_val = np.random.rand() + 0.5 # random val [0.5, 1.5]
images = [
TF.adjust_saturation(i, sat_val) if i.mode == "RGB" else i
for i in images
]
# Convert to tensor
images = [TF.to_tensor(i) for i in images]
# Normalize RGB images
images = [
TF.normalize(i, (0.5, 0.5, 0.5),
(0.5, 0.5, 0.5)) if i.shape[0] == 3 else i
for i in images
]
return images
def forward(self, image, mask):
fn_idx = torch.randperm(4)
for fn_id in fn_idx:
if (fn_id == 0) and (self.brightness is not None):
brightness = self.brightness
brightness_factor = torch.tensor(1.0).uniform_(
brightness[0], brightness[1]).item()
image = F.adjust_brightness(image, brightness_factor)
if (fn_id == 1) and (self.contrast is not None):
contrast = self.contrast
contrast_factor = torch.tensor(1.0).uniform_(
contrast[0], contrast[1]).item()
image = F.adjust_contrast(image, contrast_factor)
if (fn_id == 2) and (self.saturation is not None):
saturation = self.saturation
saturation_factor = torch.tensor(1.0).uniform_(
saturation[0], saturation[1]).item()
image = F.adjust_saturation(image, saturation_factor)
if (fn_id == 3) and (self.hue is not None):
hue = self.hue
hue_factor = torch.tensor(1.0).uniform_(hue[0], hue[1]).item()
image = F.adjust_hue(image, hue_factor)
return image, mask
def __call__(self, sample: Dict[str,np.ndarray]) -> Dict[str,np.ndarray]:
"""
Randomly adjust hue factor of the image with probablity self.p .
Args:
sample (Dict[str,np.ndarray]): image and bounding boxes to be augmented in the format
{"img":np.ndarray,"annot":np.ndarray}
Returns:
Dict[str,np.ndarray]: augmented image and bounding boxes in the format
{"img":np.ndarray,"annot":np.ndarray}
"""
if random.random() > self.p:
return sample
img,bboxes = sample["img"], sample["annot"]
img = img*255.
#print(f"In Hue: Factor={self.hue_factor}")
img = img.astype(np.uint8)
img = TF.to_tensor(img)
img = TF.to_pil_image(img,mode="RGB")
img = TF.adjust_hue(img, self.hue_factor)
img = TF.pil_to_tensor(img)
img = img.permute(1,2,0).numpy().astype(np.float32)
img = img/255.
sample = {"img": img.copy(), "annot": bboxes.copy()}
return sample
def recolor(im, recolor_net, aug=False):
N = im.shape[0]
if aug:
im_ = []
for i in range(N):
im_i = im[i]
brightness_factor = float(torch.Tensor(1).uniform_(0.8, 1.2))
contrast_factor = float(torch.Tensor(1).uniform_(0.8, 1.2))
gamma = float(torch.Tensor(1).uniform_(0.8, 1.2))
hue_factor = float(torch.Tensor(1).uniform_(-0.1, 0.1))
im_i = Ft.adjust_brightness(im_i, brightness_factor)
im_i = Ft.adjust_contrast(im_i, contrast_factor)
im_i = Ft.to_tensor(
Ft.adjust_gamma(Ft.to_pil_image(im_i.cpu()),
gamma)).cuda()
im_i = Ft.to_tensor(
Ft.adjust_hue(Ft.to_pil_image(im_i.cpu()),
hue_factor)).cuda()
im_.append(im_i)
im = torch.stack(im_, dim=0)
imout = recolor_net(im - 0.5) + im - 0.5
imout = F.sigmoid(imout * 5)
return imout
def get_params(brightness, contrast, saturation, hue):
transforms = []
if brightness > 0:
brightness_factor = random.uniform(max(0, 1 - brightness),
1 + brightness)
transforms.append(
Lambda(
lambda img: F.adjust_brightness(img, brightness_factor)))
if contrast > 0:
contrast_factor = random.uniform(max(0, 1 - contrast),
1 + contrast)
transforms.append(
Lambda(lambda img: F.adjust_contrast(img, contrast_factor)))
if saturation > 0:
saturation_factor = random.uniform(max(0, 1 - saturation),
1 + saturation)
transforms.append(
Lambda(
lambda img: F.adjust_saturation(img, saturation_factor)))
if hue > 0:
hue_factor = random.uniform(-hue, hue)
transforms.append(
Lambda(lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
transform = Compose(transforms)
return transform
def __getitem__(self,index):
if self.mode == 1:
# mode =1 为预测时使用，会从所有WSI文件中返回全部的region的图像
slideIDX = self.slideIDX[index]
coord = self.grid[index]
img = self.slides[slideIDX].read_region(coord,self.level,(self.patch_size[slideIDX],\
self.patch_size[slideIDX])).convert('RGB')
if img.size != (224,224):
img = img.resize((224,224),Image.BILINEAR)
if self.transform is not None:
img = self.transform(img)
return img
elif self.mode == 2:
# mode =2 为训练时使用，只会根据指定的index(经过上一轮MIL过程得出) \
# 从全部WSI文件中筛选对应的坐标列表,返回相应的训练图像和label
slideIDX, coord, target,h_value = self.t_data[index]
img = self.slides[slideIDX].read_region(coord,self.level,(self.patch_size[slideIDX],\
self.patch_size[slideIDX])).convert('RGB')
if h_value > 0:
hue_factor = random.uniform(h_value,0.1)
elif h_value == 0:
hue_factor = random.uniform(0,0)
elif h_value < 0:
hue_factor = random.uniform(-0.1,h_value)
img = functional.adjust_hue(img,hue_factor)
# 只有在训练模式下才进行H通道变换的颜色增强方法
# 如果在maketraindata方法设置采样复制,那么就会针对h_value的值进行不同方向的hue_factor生成,\
# 从而达到复制的样本和原来的样本有不一样的增强的效果
if img.size != (224,224):
img = img.resize((224,224),Image.BILINEAR)
if self.transform is not None:
img = self.transform(img)
return img, target
def __call__(self, img: Image.Image):
if self.i == 0:
self.params = self.get_params()
self.i = (self.i + 1) % self.len
return TF.adjust_hue(img, self.params)
def get_params(brightness, contrast, saturation, hue):
"""Get a randomized transform to be applied on image.
Arguments are same as that of __init__.
Returns:
Transform which randomly adjusts brightness, contrast and
saturation in a random order.
"""
transforms = []
if brightness > 0:
brightness_factor = random.uniform(max(0, 1 - brightness), 1 + brightness)
transforms.append(Lambda(lambda img: F.adjust_brightness(img, brightness_factor)))
if contrast > 0:
contrast_factor = random.uniform(max(0, 1 - contrast), 1 + contrast)
transforms.append(Lambda(lambda img: F.adjust_contrast(img, contrast_factor)))
if saturation > 0:
saturation_factor = random.uniform(max(0, 1 - saturation), 1 + saturation)
transforms.append(Lambda(lambda img: F.adjust_saturation(img, saturation_factor)))
if hue > 0:
hue_factor = random.uniform(-hue, hue)
transforms.append(Lambda(lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
transform = Compose(transforms)
return transform
def apply_transforms(self, video):
if not self.transform: return video
if 'train' == self.phase:
# Resize
resize = transforms.Resize(size=(self.crop_size+20, self.crop_size+20))
video = [resize(im) for im in video]
# Random crop
i, j, h, w = transforms.RandomCrop.get_params(
video[0], output_size=(self.crop_size, self.crop_size))
video = [TF.crop(im, i, j, h, w) for im in video]
# Random horizontal flipping
if random.random() > 0.5:
video = [TF.hflip(im) for im in video]
# transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0),
# Rotation
angles = [-12, -6, 0, 6, 12]
angle = random.choice(angles)
video = [TF.rotate(im, angle) for im in video]
#rotation_transform = MyRotationTransform()
hue_factor = random.uniform(-0.2,0.2)
video = [TF.adjust_hue(im, hue_factor) for im in video]
else:
resize = transforms.Resize(size=(self.crop_size, self.crop_size))
video = [resize(im) for im in video]
# Transform to tensor
return [TF.normalize(TF.to_tensor(im),self.mean, self.std) for im in video]
def __call__(self, x, params=None):
"""
Args:
x (numpy.ndarray or list): Image (H x W x C) or pose (3) or bounding box (4)
Returns:
numpy.ndarray or list: Transformed images.
"""
if params is None:
params = self.get_params(self.brightness, self.contrast, self.saturation, self.hue)
fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = params
if is_img(x): # x is an image
x = Image.fromarray(x)
for fn_id in fn_idx:
if fn_id == 0 and brightness_factor is not None:
x = F.adjust_brightness(x, brightness_factor)
elif fn_id == 1 and contrast_factor is not None:
x = F.adjust_contrast(x, contrast_factor)
elif fn_id == 2 and saturation_factor is not None:
x = F.adjust_saturation(x, saturation_factor)
elif fn_id == 3 and hue_factor is not None:
x = F.adjust_hue(x, hue_factor)
return np.array(x)
elif isinstance(x, (list, tuple)):
return [self.__call__(a, None if self.randomize_per_image else params) for a in x]
return x
def randomize_parameters(self):
self.p = random.random()
brightness_factor = random.uniform(self.brightness[0],
self.brightness[1])
contrast_factor = random.uniform(self.contrast[0], self.contrast[1])
saturation_factor = random.uniform(self.saturation[0],
self.saturation[1])
hue_factor = random.uniform(self.hue[0], self.hue[1])
transforms = []
transforms.append(
torchvision.transforms.Lambda(
lambda img: F.adjust_brightness(img, brightness_factor)))
transforms.append(
torchvision.transforms.Lambda(
lambda img: F.adjust_contrast(img, contrast_factor)))
transforms.append(
torchvision.transforms.Lambda(
lambda img: F.adjust_saturation(img, saturation_factor)))
transforms.append(
torchvision.transforms.Lambda(
lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
self.transform = torchvision.transforms.Compose(transforms)
def get_params(brightness, contrast, saturation, hue):
"""Get a randomized transform to be applied on image.
Arguments are same as that of __init__.
Returns:
Transform which randomly adjusts brightness, contrast and
saturation in a random order.
"""
transforms = []
if brightness is not None:
brightness_factor = random.uniform(brightness[0], brightness[1])
transforms.append(torchvision.transforms.Lambda(lambda img: F.adjust_brightness(img, brightness_factor)))
if contrast is not None:
contrast_factor = random.uniform(contrast[0], contrast[1])
transforms.append(torchvision.transforms.Lambda(lambda img: F.adjust_contrast(img, contrast_factor)))
if saturation is not None:
saturation_factor = random.uniform(saturation[0], saturation[1])
transforms.append(torchvision.transforms.Lambda(lambda img: F.adjust_saturation(img, saturation_factor)))
if hue is not None:
hue_factor = random.uniform(hue[0], hue[1])
transforms.append(torchvision.transforms.Lambda(lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
transform = torchvision.transforms.Compose(transforms)
return transform
def forward(self, img):
"""
Args:
img (PIL Image or Tensor): Input image.
Returns:
PIL Image or Tensor: Color jittered image.
"""
#Check if image is bright enough
#code form https://stackoverflow.com/questions/3490727/what-are-some-methods-to-analyze-image-brightness-using-python
im = img.convert('L')
stat = ImageStat.Stat(im)
#print(stat.rms[0])
fn_idx, brightness_factor, contrast_factor, saturation_factor, hue_factor = \
self.get_params(self.brightness, self.contrast, self.saturation, self.hue)
for fn_id in fn_idx:
if fn_id == 0 and brightness_factor is not None and (
stat.rms[0] >= 50 or brightness_factor >= 1):
img = functional.adjust_brightness(img, brightness_factor)
elif fn_id == 1 and contrast_factor is not None and (
stat.rms[0] >= 50 or contrast_factor >= 1):
img = functional.adjust_contrast(img, contrast_factor)
elif fn_id == 2 and saturation_factor is not None:
img = functional.adjust_saturation(img, saturation_factor)
elif fn_id == 3 and hue_factor is not None:
img = functional.adjust_hue(img, hue_factor)
return img
def __call__(self, image): if random.random() < 0.5: adjust_factor = random.uniform( -self.delta/255. , self.delta/255. ) # Transformation image = FT.adjust_hue(image, adjust_factor) return image
def augmentations(img):
crop_h = 32
crop_w = 128
i = torch.randint(height - crop_h + 1, size=(1, )).item()
j = torch.randint(width - crop_w + 1, size=(1, )).item()
img = F2.crop(img, i, j, crop_h, crop_w)
img = F2.resize(img, (height, width))
fn_idx = torch.randperm(4)
for fn_id in fn_idx:
if fn_id == 0 and torch.rand(1) < 0.2:
brg = float(torch.empty(1).uniform_(0.6, 2))
img = F2.adjust_brightness(img, brg)
elif fn_id == 1 and torch.rand(1) < 0.2:
con = float(torch.empty(1).uniform_(0.6, 2))
img = F2.adjust_contrast(img, con)
elif fn_id == 2 and torch.rand(1) < 0.2:
sat = float(torch.empty(1).uniform_(0.6, 2))
img = F2.adjust_saturation(img, sat)
elif fn_id == 3 and torch.rand(1) < 0.2:
hue = float(torch.empty(1).uniform_(-0.5, 0.5))
img = F2.adjust_hue(img, hue)
if torch.rand(1) < 0.2:
img = F2.hflip(img)
if torch.rand(1) < 0.2:
img = F2.vflip(img)
return img
def forward(self, img):
brightness_factor, contrast_factor, saturation_factor, hue_factor = 0, 0, 0, 0
fn_idx = torch.randperm(4)
for fn_id in fn_idx:
if fn_id == 0 and self.brightness is not None:
brightness = self.brightness
brightness_factor = torch.tensor(1.0).uniform_(brightness[0], brightness[1]).item()
img = F.adjust_brightness(img, brightness_factor)
if fn_id == 1 and self.contrast is not None:
contrast = self.contrast
contrast_factor = torch.tensor(1.0).uniform_(contrast[0], contrast[1]).item()
img = F.adjust_contrast(img, contrast_factor)
if fn_id == 2 and self.saturation is not None:
saturation = self.saturation
saturation_factor = torch.tensor(1.0).uniform_(saturation[0], saturation[1]).item()
img = F.adjust_saturation(img, saturation_factor)
if fn_id == 3 and self.hue is not None:
hue = self.hue
hue_factor = torch.tensor(1.0).uniform_(hue[0], hue[1]).item()
img = F.adjust_hue(img, hue_factor)
return img, [brightness_factor, contrast_factor, saturation_factor, hue_factor]
def randomize_parameters(self):
transforms = []
if self.brightness is not None:
brightness_factor = random.uniform(self.brightness[0],
self.brightness[1])
transforms.append(
Lambda(
lambda img: F.adjust_brightness(img, brightness_factor)))
if self.contrast is not None:
contrast_factor = random.uniform(self.contrast[0],
self.contrast[1])
transforms.append(
Lambda(lambda img: F.adjust_contrast(img, contrast_factor)))
if self.saturation is not None:
saturation_factor = random.uniform(self.saturation[0],
self.saturation[1])
transforms.append(
Lambda(
lambda img: F.adjust_saturation(img, saturation_factor)))
if self.hue is not None:
hue_factor = random.uniform(self.hue[0], self.hue[1])
transforms.append(
Lambda(lambda img: F.adjust_hue(img, hue_factor)))
random.shuffle(transforms)
self.transform = Compose(transforms)
def test_adjust_hue(self):
x_shape = [2, 2, 3]
x_data = [0, 5, 13, 54, 135, 226, 37, 8, 234, 90, 255, 1]
x_np = np.array(x_data, dtype=np.uint8).reshape(x_shape)
x_pil = Image.fromarray(x_np, mode='RGB')
with self.assertRaises(ValueError):
F.adjust_hue(x_pil, -0.7)
F.adjust_hue(x_pil, 1)
# test 0: almost same as x_data but not exact.
# probably because hsv <-> rgb floating point ops
y_pil = F.adjust_hue(x_pil, 0)
y_np = np.array(y_pil)
y_ans = [0, 5, 13, 54, 139, 226, 35, 8, 234, 91, 255, 1]
y_ans = np.array(y_ans, dtype=np.uint8).reshape(x_shape)
assert np.allclose(y_np, y_ans)
# test 1
y_pil = F.adjust_hue(x_pil, 0.25)
y_np = np.array(y_pil)
y_ans = [13, 0, 12, 224, 54, 226, 234, 8, 99, 1, 222, 255]
y_ans = np.array(y_ans, dtype=np.uint8).reshape(x_shape)
assert np.allclose(y_np, y_ans)
# test 2
y_pil = F.adjust_hue(x_pil, -0.25)
y_np = np.array(y_pil)
y_ans = [0, 13, 2, 54, 226, 58, 8, 234, 152, 255, 43, 1]
y_ans = np.array(y_ans, dtype=np.uint8).reshape(x_shape)
assert np.allclose(y_np, y_ans)
def test_adjusts_L_mode(self):
x_shape = [2, 2, 3]
x_data = [0, 5, 13, 54, 135, 226, 37, 8, 234, 90, 255, 1]
x_np = np.array(x_data, dtype=np.uint8).reshape(x_shape)
x_rgb = Image.fromarray(x_np, mode='RGB')
x_l = x_rgb.convert('L')
assert F.adjust_brightness(x_l, 2).mode == 'L'
assert F.adjust_saturation(x_l, 2).mode == 'L'
assert F.adjust_contrast(x_l, 2).mode == 'L'
assert F.adjust_hue(x_l, 0.4).mode == 'L'
assert F.adjust_gamma(x_l, 0.5).mode == 'L'
def torchvision(self, img):
img = torchvision.adjust_hue(img, hue_factor=0.1)
img = torchvision.adjust_saturation(img, saturation_factor=1.2)
img = torchvision.adjust_brightness(img, brightness_factor=1.2)
return img