Transformations of pytorch torch torch vision

Partial reference ：
PyTorch Learning notes （ 3、 ... and ）：transforms Twenty two ways to
pytorch torchvision transform

1. Yes PIL.Image To transform

1.1. Combine

transform：class torchvision.transforms.Compose(transforms)
Will be multiple transform Use in combination .
transforms： from transform The list of components . Example ：

transforms.Compose([
     transforms.CenterCrop(10),
     transforms.ToTensor(),
 ])

1.2. Reshape size

class torchvision.transforms.Scale(size, interpolation=2)
Will input PIL.Image Resize to a given size size,size Is the side length of the smallest side . for instance , If the original height>width, So the size of the picture after changing the size is (size*height/width, size).
Use cases :

from torchvision import transforms
from PIL import Image
crop = transforms.Scale(12)
img = Image.open('test.jpg')

print(type(img))
print(img.size)

croped_img=crop(img)
print(type(croped_img))
print(croped_img.size)

<class 'PIL.PngImagePlugin.PngImageFile'>
(10, 10)
<class 'PIL.Image.Image'>
(12, 12)

1.3. Center cut

class torchvision.transforms.CenterCrop(size)
Will be given PIL.Image Make a center cut , Get the given size,size It can be tuple,(target_height, target_width).size It can also be a Integer, under these circumstances , The shape of the cut image is square .

1.4. Random center cut

class torchvision.transforms.RandomCrop(size, padding=0)
The position of the cutting center point is randomly selected .size It can be tuple It can also be Integer.

1.5. Random horizontal flip

class torchvision.transforms.RandomHorizontalFlip(p=0.5)
Randomly flip a given... Horizontally PIL.Image, The probability of 0.5. namely ： Half the chance to flip , Half the probability doesn't flip .

1.6. Cut at random ＋ Reshape size

class torchvision.transforms.RandomSizedCrop(size, interpolation=2)
First, the given PIL.Image Random cut , And then again resize Into a given size size .

1.7. fill

class torchvision.transforms.Pad(padding, fill=0)
Will be given PIL.Image All the edges of a given pad value fill .

• padding： How many pixels to fill
• fill： What values to fill in
  Example ：

from torchvision import transforms
from PIL import Image
padding_img = transforms.Pad(padding=10, fill=0)
img = Image.open('test.jpg')

print(type(img))
print(img.size)

padded_img=padding(img)
print(type(padded_img))
print(padded_img.size)

<class 'PIL.PngImagePlugin.PngImageFile'>
(10, 10)
<class 'PIL.Image.Image'>
(30, 30) # Because it has to be filled up, down, left and right 10 Pixel , So after filling size yes (30,30)

1.8. Random vertical flip

class torchvision.transforms.RandomVerticalFlip(p=0.5)
Randomly flip the given... Vertically PIL.Image, The probability of 0.5. namely ： Half the chance to flip , Half the probability doesn't flip .

1.9. Random rotation

class torchvision.transforms.Rotation(degrees, resample=False, expand=False, center=None)
In accordance with the degrees Randomly rotate a certain angle

• degress：(sequence or float or int) , If it is a single number , Such as 30, It means in （-30,+30） Random rotation between
  if sequence, Such as (30,60), It means in 30-60 Rotate randomly between degrees
• resample： Resampling method selection , Optional PIL.Image.NEAREST, PIL.Image.BILINEAR, PIL.Image.BICUBIC, The default is nearest neighbor
• center: You can select center rotation or upper left rotation

2. Yes Tensor To transform

2.1. Standardization

class torchvision.transforms.Normalize(mean, std)
Given the mean ：(R,G,B) variance ：（R,G,B）, It will put Tensor Regularization . namely ：Normalized_image=(image-mean)/std.

3.Conversion Transforms

3.1. Convert to tensor

class torchvision.transforms.ToTensor
Let's take a value range [0,255] Of PIL.Image perhaps shape by (H,W,C) Of numpy.ndarray, Convert to shape [C,H,W], The value range is [0,1.0] Of torch.FloadTensor

data = np.random.randint(0, 255, size=300)
img = data.reshape(10,10,3)
print(img.shape)
img_tensor = transforms.ToTensor()(img) #  convert to tensor
print(img_tensor)

3.2. Convert to PIL image

class torchvision.transforms.ToPILImage
take shape by (C,H,W) Of Tensor or shape by (H,W,C) Of numpy.ndarray convert to PIL.Image, The value remains the same .

4. Universal transformation

4.1.class torchvision.transforms.Lambda(lambd)

Use lambd As a converter .