def make_transform(sz_resize=256,
sz_crop=227,
mean=None,
std=None,
rgb_to_bgr=True,
is_train=True,
intensity_scale=None,
rotate=0,
jitter=0,
hue=0):
if std is None:
std = [1, 1, 1]
if mean is None:
mean = [104, 117, 128]
return transforms.Compose([
RGBToBGR() if rgb_to_bgr else Identity(),
transforms.RandomRotation(rotate) if is_train else Identity(),
transforms.RandomResizedCrop(sz_crop) if is_train else Identity(),
transforms.Resize(sz_resize) if not is_train else Identity(),
transforms.CenterCrop(sz_crop) if not is_train else Identity(),
transforms.RandomHorizontalFlip() if is_train else Identity(),
#transforms.ColorJitter(brightness=jitter, contrast=jitter, saturation=jitter, hue=hue) if is_train else Identity(),
#### New Added ####
transforms.RandomPerspective() if is_train else Identity(),
transforms.RandomEqualize() if is_train else Identity(),
transforms.RandomVerticalFlip() if is_train else Identity(),
transforms.RandomAdjustSharpness(
sharpness_factor=2) if is_train else Identity(),
transforms.RandomAutocontrast() if is_train else Identity(),
transforms.RandomErasing() if is_train else Identity(),
transforms.RandomAffine(degrees=(-60, 60),
translate=(0.1, 0.3),
scale=(0.5, 0.75)) if is_train else Identity(),
#### New Added ####
transforms.ToTensor(),
ScaleIntensities(
*intensity_scale) if intensity_scale is not None else Identity(),
transforms.Normalize(
mean=mean,
std=std,
)
])
# RandomAutocontrast
# ~~~~~~~~~~~~~~~~~~
# The :class:`~torchvision.transforms.RandomAutocontrast` transform
# (see also :func:`~torchvision.transforms.functional.autocontrast`)
# randomly applies autocontrast to the given image.
autocontraster = T.RandomAutocontrast()
autocontrasted_imgs = [autocontraster(orig_img) for _ in range(4)]
plot(autocontrasted_imgs)
####################################
# RandomEqualize
# ~~~~~~~~~~~~~~
# The :class:`~torchvision.transforms.RandomEqualize` transform
# (see also :func:`~torchvision.transforms.functional.equalize`)
# randomly equalizes the histogram of the given image.
equalizer = T.RandomEqualize()
equalized_imgs = [equalizer(orig_img) for _ in range(4)]
plot(equalized_imgs)
####################################
# AutoAugment
# ~~~~~~~~~~~
# The :class:`~torchvision.transforms.AutoAugment` transform
# automatically augments data based on a given auto-augmentation policy.
# See :class:`~torchvision.transforms.AutoAugmentPolicy` for the available policies.
policies = [T.AutoAugmentPolicy.CIFAR10, T.AutoAugmentPolicy.IMAGENET, T.AutoAugmentPolicy.SVHN]
augmenters = [T.AutoAugment(policy) for policy in policies]
imgs = [
[augmenter(orig_img) for _ in range(4)]
for augmenter in augmenters
]
from matplotlib.image import AxesImage
from hearth.vision.datasets import RGBImageDataset
from hearth.vision.transforms import RandomFlipOrient
_ids = [str(uuid.uuid4()) for i in range(100)]
_augmentations = T.RandomApply(
[
T.RandomHorizontalFlip(p=0.5),
RandomFlipOrient(0.5),
T.ColorJitter(brightness=0.1, hue=0.1),
T.RandomAdjustSharpness(sharpness_factor=2),
T.RandomAutocontrast(),
T.RandomEqualize(),
],
p=0.8,
)
@pytest.fixture(scope='module')
def image_dir(tmpdir_factory):
d = tmpdir_factory.mktemp('images')
dirname = str(d)
# compute and save 100 random images of random sizes
for _id in _ids:
w = random.randint(32, 512)
h = random.randint(32, 512)
img = torch.rand(3, w, h)
def data_augmentation(ToTensor=False,
Resize=None,
Contrast=None,
Equalize=None,
HFlip=None,
Invert=None,
VFlip=None,
Rotation=None,
Grayscale=None,
Perspective=None,
Erasing=None,
Crop=None):
'''
DataAgumentation 2021/03/23 by Mr.w
-------------------------------------------------------------
ToTensor : False/True , 注意转为Tensor，通道会放在第一维
Resize : tuple-->(500,500)
Contrast : 0-1 -->图像被自动对比度的可能
Equalize : 0-1 -->图像均衡可能性
HFlip : 0-1 --> 图像水平翻转
Invert : 0-1--> 随机翻转
VFlip : 0-1 --> 图像垂直翻转
Rotation : 0-360 --> 随机旋转度数范围, as : 90 , [-90,90]
Grayscale : 0-1 --> 随机转换为灰度图像
Perspective : 0-1 --> 随机扭曲图像
Erasing : 0-1 --> 随机擦除
Crop : tuple --> (500,500)
-------------------------------------------------------------
return : transforms.Compose(train_transform) --> 方法汇总
'''
#列表导入Compose
train_transform = []
if ToTensor == True:
trans_totensor = transforms.ToTensor()
train_transform.append(trans_totensor)
if Resize != None:
trans_Rsize = transforms.Resize(Resize) # Resize=(500,500)
train_transform.append(trans_Rsize)
if Contrast != None:
trans_Rcontrast = transforms.RandomAutocontrast(p=Contrast)
train_transform.append(trans_Rcontrast)
if Equalize != None:
trans_REqualize = transforms.RandomEqualize(p=Equalize)
train_transform.append(trans_REqualize)
if HFlip != None:
train_transform.append(transforms.RandomHorizontalFlip(p=HFlip))
if Invert != None:
train_transform.append(transforms.RandomInvert(p=Invert))
if VFlip != None:
train_transform.append(transforms.RandomVerticalFlip(p=VFlip))
if Rotation != None:
train_transform.append(
transforms.RandomRotation(Rotation,
expand=False,
center=None,
fill=0,
resample=None))
if Grayscale != None:
train_transform.append(transforms.RandomGrayscale(p=Grayscale))
if Perspective != None:
train_transform.append(
transforms.RandomPerspective(distortion_scale=0.5,
p=Perspective,
fill=0))
if Erasing != None:
train_transform.append(
transforms.RandomErasing(p=Erasing,
scale=(0.02, 0.33),
ratio=(0.3, 3.3),
value=0,
inplace=False))
if Crop != None:
train_transform.append(
transforms.RandomCrop(Crop,
padding=None,
pad_if_needed=False,
fill=0,
padding_mode='constant'))
return transforms.Compose(train_transform)