2.2 【pytorch】torchvision. transforms

About torchvision and transforms

（1）torchvision There are four functional modules ：model、dataset、transforms、utils
（2）transforms modular The source data can be preprocessed 、 enhance
（3）transforms Provide for the right to PIL Image Objects and Tensor Common operations of objects . It can be done by PIL Read pictures directly , PIL.Image.open(path) ;

Official document address ： https://pytorch.org/vision/stable/transforms.html

transforms

1、 tailoring （Crop）

Center cut ：transforms.CenterCrop

torchvision.transforms.CenterCrop(size)

Parameters ：
size: The size of the image to be cropped

from PIL import Image
import matplotlib.pyplot as plt
import torchvision.transforms as transforms

img_src = Image.open('./bird.jpg')

img_1 = transforms.CenterCrop(200)(img_src)
img_2 = transforms.CenterCrop((200, 200))(img_src)
img_3 = transforms.CenterCrop((300, 200))(img_src)
img_4 = transforms.CenterCrop((500, 500))(img_src)

plt.subplot(231)
plt.imshow(img_src)
plt.subplot(232)
plt.imshow(img_1)
plt.subplot(233)
plt.imshow(img_2)
plt.subplot(234)
plt.imshow(img_3)
plt.subplot(235)
plt.imshow(img_4)
plt.show()

We can see from the above examples ：
（1） If you cut a square ,transforms.CenterCrop(100) and transforms.CenterCrop((100, 100)), Two kinds of writing size Methods , The effect is the same
（2） If the set output picture size is larger than the original size , Will fill the edges with black

---

Random cutting ：transforms.RandomCrop

#  According to the given size Random cutting 
torchvision.transforms.RandomCrop(size, 
                      padding = None, 
                      pad_if_needed = False, 
                      fill=0, 
                      padding_mode ='constant')

function ：
Randomly cut out the size of... From the picture size Pictures of the , If there is padding, Then go ahead padding, Then cut randomly size Size picture .

Parameters ：
size
padding: Set fill size
– – When it comes to a when , Fill up, down, left and right a Pixel
– – When it comes to (a, b) when , Fill left and right a Pixel , Fill it up and down b Pixel
– – When it comes to (a, b, c, d) when , Fill the upper left and the lower right respectively a,b,c,d
pad_if_needed： When the picture is smaller than the set size, Whether to fill in
padding_mode：
– – constant: The pixel value is determined by fill Set up （ Default ）
– – edge: The pixel value is set by the image edge pixels
– – reflect: Mirror fill , The last pixel is not mirrored .([1,2,3,4] -> [3,2,1,2,3,4,3,2])
– – symmetric: Mirror fill , The last pixel is also mirrored .([1,2,3,4] -> [2,1,1,2,3,4,4,4,3])
fill： When padding_mode by constant when , Sets the pixel value of the fill （ The default is 0）

Random aspect ratio tailoring ：transforms.RandomResizedCrop `

torchvision.transforms.RandomResizedCrop(size, 
                                         scale=(0.08, 1.0), 
                                         ratio=(0.75, 1.3333333333333333), 
                                         interpolation=2)

function ：
The random size 、 Crop pictures with random aspect ratio . First of all, according to the scale Cut the original drawing to the scale of , And then according to ratio And then cut the length to width ratio , Finally, the image is transformed into size size .

Parameters ：
size: Cropped image size
scale: Randomly scale the area , Default random selection (0.08, 1) A number between
ratio: Random aspect ratio , Default random selection ($\frac{3}{4}$, $\frac{4}{3}$ ) A number between . Because beyond this ratio, there will be obvious distortion
interpolation: When the cropped image is smaller than size when , We need to use interpolation method resize
– – PIL.Image.NEAREST
– – PIL.Image.BILINEAR
– – PIL.Image.BICUBIC

from PIL import Image
import matplotlib.pyplot as plt
import torchvision.transforms as transforms

img_src = Image.open('./bird.jpg')

img_6 = transforms.RandomResizedCrop((200, 300), scale=(1, 1))(img_src)
img_7 = transforms.RandomResizedCrop((200, 300), scale=(0.5, 0.5))(img_src)
img_8 = transforms.RandomResizedCrop((200, 300), scale=(1, 1),  ratio=(0.5, 0.5))(img_src)
img_9 = transforms.RandomResizedCrop((400, 600), scale=(1, 1),  ratio=(0.5, 0.5))(img_src)

plt.subplot(231)
plt.imshow(img_src)
plt.subplot(232)
plt.imshow(img_6)
plt.subplot(233)
plt.imshow(img_7)
plt.subplot(234)
plt.imshow(img_8)
plt.subplot(235)
plt.imshow(img_9)
plt.show()

Top, bottom, left and right center cut ：transforms.FiveCrop

Cut the top, bottom, left and right center, and then flip ,transforms.TenCrop

2、 Flip and rotate （Flip and Rotation）

According to probability p Flip horizontal ：transforms.RandomHorizontalFlip(p=0.5)

According to probability p Flip vertically ：transforms.RandomVerticalFlip(p=0.5)

Random rotation ：transforms.RandomRotation

---

3、 Image transformation

• transforms.Resize
• Standardization ：transforms.Normalize
• To tensor, And normalized to [0-1]：transforms.ToTensor
• fill ：transforms.Pad
• Change the brightness 、 Contrast and saturation ：transforms.ColorJitter
• Go to grayscale ：transforms.Grayscale
• linear transformation ：transforms.LinearTransformation()
• Affine transformation ：transforms.RandomAffine
• According to probability p Turn to grayscale ：transforms.RandomGrayscale
• Convert data to PILImage：transforms.ToPILImage
• transforms.Lambda：Apply a user-defined lambda as a transform

---

4、transforms Method of operation

• Make data enhancement more flexible transforms.RandomChoice(transforms), From a given series of transforms Choose one of them to operate
• transforms.RandomApply(transforms, p=0.5), Give me a transform Plus the probability , Operate according to probability
• transforms.RandomOrder, take transforms The operations in are randomly disrupted

---

reference：

https://www.cnblogs.com/zhangxiann/p/13570884.html

Official document address ： https://pytorch.org/vision/stable/transforms.html