【pyTorch基础】数据加载和处理

from PIL import Image
from torchvision import transforms as T
import torch

to_tensor = T.ToTensor() # 转换一个PIL库的图片或者numpy的数组转为tensor张量类型:[0,255]->[0,1]
to_pil = T.ToPILImage() # 将Numpy的ndarray或者Tensor转化成PILImage类型
cat = Image.open('./data/image/cat.jpg')

cat_c = to_tensor(cat) # 图片转张量
print(cat_c.shape)
cat_c = to_pil(cat_c) # 张量转PILImage

torch.Size([3, 1274, 1280])

# 将tensor转换为PILImage
t_out = torch.randn(3,256,256)
img1 = T.ToPILImage()(t_out)
img1

torchvision.transforms.CenterCrop:从图像中心裁剪图片

transforms = T.Compose([T.Resize(224),T.CenterCrop(224),T.ToTensor()])# 当T.CenterCrop()的参数大于T.Resize()的参数时，周围用0(黑色)填充
print(transforms)
# 先将整张图片大小缩小为长宽224，再对中心裁剪长宽224，相当于缩小图片
# T.Compose:将操作整合
# 生成了一个带有T.resize等功能的转换器
cat_t = transforms(cat)# 将图片放入转换器进行转换
print(cat_t.shape)
to_pil(cat_t) # 将tensor格式的图片转为pil

Compose(
    Resize(size=224, interpolation=bilinear, max_size=None, antialias=None)
    CenterCrop(size=(224, 224))
    ToTensor()
)
torch.Size([3, 224, 224])

torchvision.transforms.RandomCrop:从图片中随机裁剪出尺寸为size的图片

transforms = T.Compose([T.Resize(224),T.RandomCrop(256, padding=(16, 64)),T.ToTensor()])
# 先上下填充64，左右填充16(图像变大了),再随机256*256裁剪
cat_t = transforms(cat)
print(cat_t.shape)  # 3*224*224
to_pil(cat_t)

torch.Size([3, 256, 256])

# padding_mode:constant,edge,reflect,symmetric，constant时像素值由fill设定
transforms = T.Compose([T.Resize(224), T.RandomCrop(352,padding=(64,64),padding_mode = 'constant', fill=66),T.ToTensor()])
cat_t = transforms(cat)
to_pil(cat_t)

# torchvision.transforms.RandomHorizontalFlip
transforms = T.Compose([T.Resize(224),T.RandomHorizontalFlip(p=0.5),T.ToTensor()])  # 以0.5的概率左右翻转
cat_t = transforms(cat)  
cat_t.shape 
to_pil(cat_t)

# torchvision.transforms.RandomRotation
transforms = T.Compose([T.Resize(224),T.RandomRotation(30, center=(0, 0), expand=True),T.ToTensor()]) 
cat_t = transforms(cat)  
cat_t.shape  
to_pil(cat_t)

transform方法

裁剪

1. transforms.CenterCrop
2. transforms.RandomCrop
3. transforms.RandomResizedCrop
4. transforms.FiveCrop
5. transforms.TenCrop

翻转和旋转

1. transforms.RandomHorizontalFlip
2. transforms.RandomVerticalFlip
3. transforms.RandomRotation

图像变换

1. transforms.Pad
2. transforms.ColorJitter
3. transforms.Grayscale
4. transforms.RandomGrayscale
5. transforms.RandomAffine
6. transforms.LinearTransformation
7. transforms.RandomErasing
8. transforms.Lambda
9. transforms.Resize
10. transforms.Totensor
11. transforms.Normalize

transforms的操作

1. transforms.RandomChoice
2. transforms.RandomApply
3. transforms.RandomOrder