torchvision 笔记：transforms.Normalize()

Posted 2021-10-22 UQI-LIUWJ

        一般和transforms.ToTensor()搭配使用

        作用就是先将输入归一化到(0,1)【transforms.ToTensor()】，再使用公式"(x-mean)/std"，将每个元素分布到(-1,1)

   很多CV的代码中，是这样使用这一条语句的：

torchvision.transforms.Normalize(
    mean=[0.485, 0.456, 0.406], 
    std=[0.229, 0.224, 0.225])

这一组参数是从ImageNet数据集中获得的

在 torchvision 笔记：ToTensor()_UQI-LIUWJ的博客-CSDN博客的代码基础上我们进行修改

ToTensor 中的代码：

from PIL import Image
from torchvision import transforms, utils
a=Image.open(b+'img/00000.jpg')
a

 

y=transforms.ToTensor()
a=y(a)
a
'''
tensor([[[0.9255, 0.9255, 0.9255,  ..., 0.9176, 0.9176, 0.9176],
         [0.9255, 0.9255, 0.9255,  ..., 0.9176, 0.9176, 0.9176],
         [0.9255, 0.9255, 0.9255,  ..., 0.9176, 0.9176, 0.9176],
         ...,
         [0.7882, 0.7882, 0.7882,  ..., 0.7922, 0.7922, 0.7922],
         [0.7882, 0.7882, 0.7882,  ..., 0.7922, 0.7922, 0.7922],
         [0.7882, 0.7882, 0.7882,  ..., 0.7922, 0.7922, 0.7922]],

        [[0.9255, 0.9255, 0.9255,  ..., 0.9216, 0.9216, 0.9216],
         [0.9255, 0.9255, 0.9255,  ..., 0.9216, 0.9216, 0.9216],
         [0.9255, 0.9255, 0.9255,  ..., 0.9216, 0.9216, 0.9216],
         ...,
         [0.7961, 0.7961, 0.7961,  ..., 0.7922, 0.7922, 0.7922],
         [0.7961, 0.7961, 0.7961,  ..., 0.7922, 0.7922, 0.7922],
         [0.7961, 0.7961, 0.7961,  ..., 0.7922, 0.7922, 0.7922]],

        [[0.9255, 0.9255, 0.9255,  ..., 0.9294, 0.9294, 0.9294],
         [0.9255, 0.9255, 0.9255,  ..., 0.9294, 0.9294, 0.9294],
         [0.9255, 0.9255, 0.9255,  ..., 0.9294, 0.9294, 0.9294],
         ...,
         [0.7922, 0.7922, 0.7922,  ..., 0.8000, 0.8000, 0.8000],
         [0.7922, 0.7922, 0.7922,  ..., 0.8000, 0.8000, 0.8000],
         [0.7922, 0.7922, 0.7922,  ..., 0.8000, 0.8000, 0.8000]]])
'''

 Normalize的代码

z=transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
a=z(a)
a
'''
tensor([[[1.9235, 1.9235, 1.9235,  ..., 1.8893, 1.8893, 1.8893],
         [1.9235, 1.9235, 1.9235,  ..., 1.8893, 1.8893, 1.8893],
         [1.9235, 1.9235, 1.9235,  ..., 1.8893, 1.8893, 1.8893],
         ...,
         [1.3242, 1.3242, 1.3242,  ..., 1.3413, 1.3413, 1.3413],
         [1.3242, 1.3242, 1.3242,  ..., 1.3413, 1.3413, 1.3413],
         [1.3242, 1.3242, 1.3242,  ..., 1.3413, 1.3413, 1.3413]],

        [[2.0959, 2.0959, 2.0959,  ..., 2.0784, 2.0784, 2.0784],
         [2.0959, 2.0959, 2.0959,  ..., 2.0784, 2.0784, 2.0784],
         [2.0959, 2.0959, 2.0959,  ..., 2.0784, 2.0784, 2.0784],
         ...,
         [1.5182, 1.5182, 1.5182,  ..., 1.5007, 1.5007, 1.5007],
         [1.5182, 1.5182, 1.5182,  ..., 1.5007, 1.5007, 1.5007],
         [1.5182, 1.5182, 1.5182,  ..., 1.5007, 1.5007, 1.5007]],

        [[2.3088, 2.3088, 2.3088,  ..., 2.3263, 2.3263, 2.3263],
         [2.3088, 2.3088, 2.3088,  ..., 2.3263, 2.3263, 2.3263],
         [2.3088, 2.3088, 2.3088,  ..., 2.3263, 2.3263, 2.3263],
         ...,
         [1.7163, 1.7163, 1.7163,  ..., 1.7511, 1.7511, 1.7511],
         [1.7163, 1.7163, 1.7163,  ..., 1.7511, 1.7511, 1.7511],
         [1.7163, 1.7163, 1.7163,  ..., 1.7511, 1.7511, 1.7511]]])
'''

将tensor反变换回图片，则有