PIL Image

 
from PIL import Image
img = Image.open("lingmao.jpg")

 
type(img)  #PIL.JpegImagePlugin.JpegImageFile

 
img.size  #(940, 940)

transforms.Resize

 
import torch
from torchvision import transforms

img = transforms.Resize(size=(256, 256))(img)

 
import numpy as np

np.array(img).shape  #(256, 256, 3)

transforms.Grayscale

 
img = transforms.Grayscale()(img)

 
np.array(img)

array([[ 85,  86,  87, ..., 100, 100, 100],
       [ 86,  87,  87, ..., 100, 100, 100],
       [ 87,  87,  87, ..., 100, 100, 100],
       ...,
       [ 38,  38,  38, ...,  40,  39,  38],
       [ 36,  36,  36, ...,  39,  38,  38],
       [ 36,  36,  36, ...,  38,  38,  38]], dtype=uint8)

transforms.ToTensor

转tensor且转换到[0,1] 之间

 
transforms.ToTensor()(img)

 
tensor([[[0.3333, 0.3373, 0.3412,  ..., 0.3922, 0.3922, 0.3922],
    [0.3373, 0.3412, 0.3412,  ..., 0.3922, 0.3922, 0.3922],
    [0.3412, 0.3412, 0.3412,  ..., 0.3922, 0.3922, 0.3922],
    ...,
    [0.1490, 0.1490, 0.1490,  ..., 0.1569, 0.1529, 0.1490],
    [0.1412, 0.1412, 0.1412,  ..., 0.1529, 0.1490, 0.1490],
    [0.1412, 0.1412, 0.1412,  ..., 0.1490, 0.1490, 0.1490]]])

直接除以255，再取小数点后4位有效数字

 
100/255 = 0.39215686274509803

38/255  = 0.14901960784313725

transforms.Normalize

 
img = transforms.ToTensor()(img)

type(img) # torch.Tensor

按特征维度进行归一化，即批归一化

 
#output[channel] = (input[channel] - mean[channel]) / std[channel]

transforms.Normalize(mean=[0.5], std=[0.5])(img)

 
tensor([[[-0.3333, -0.3255, -0.3176,  ..., -0.2157, -0.2157, -0.2157],
    [-0.3255, -0.3176, -0.3176,  ..., -0.2157, -0.2157, -0.2157],
    [-0.3176, -0.3176, -0.3176,  ..., -0.2157, -0.2157, -0.2157],
    ...,
    [-0.7020, -0.7020, -0.7020,  ..., -0.6863, -0.6941, -0.7020],
    [-0.7176, -0.7176, -0.7176,  ..., -0.6941, -0.7020, -0.7020],
    [-0.7176, -0.7176, -0.7176,  ..., -0.7020, -0.7020, -0.7020]]])

参考