torchvision 中transform的使用

torchvision 中transform 的使用

torchvision.transforms方法

• 把PIL.Image转换成torch.FloatTensor

• 把shape为（H,W,C)的numpy.ndarry转换成(C,H,W),取值范围是[0,1.0]的torch.FloatTensor

  • (H,W,C)以为(高，宽，通道数)

  • 黑白图片的通道数为1，每个像素点的取值为[0,255]

  • 彩色图片的通道数为(R,G,B)，每个通道的每个像素点的取值为[0,255]

  transforms.ToTensor()(img)

  • tensor.permute(2,0,1)

import torchvision
img_pil=mnist[0][0]
img_torch=torchvision.transforms.ToTensor()(img_pil)
import matplotlib.pyplot as plt
%matplotlib inline
print('img_pil.size = ',img_pil.size)
print('img_torch.shape = ',img_torch.shape)
print(img_torch)
plt.imshow(img_pil)

img_pil.size =  (28, 28)
img_torch.shape =  torch.Size([1, 28, 28])
41
<matplotlib.image.AxesImage at 0x1fb1fa24d68>

img_pil_1=torchvision.transforms.ToPILImage()(img_torch)
print(img_pil_1.size)
plt.imshow(img_pil_1)

(28, 28)
38
<matplotlib.image.AxesImage at 0x1fb1f9c5358>

#tensor到numpy的转化
import numpy as np
img_np=np.array(img_torch.permute(1,2,0))
print(img_np.shape)

(28, 28, 1)

torchvision.tranforms.Normalize(mean,std)
$$z=\frac{x-mean}{std}$$
给定均值：mean, shape和图片的通道数相同，方差：std，和图片的通道数相同，将会把Tensor规范化处理。

a=transforms.ToTensor()(a)
normal=transforms.Normalize([1,1,1],[1,1,1])
print(normal)
c=normal(a)
print(c)

Normalize(mean=[1, 1, 1], std=[1, 1, 1])
tensor([[[-0.1059, -0.1059, -0.1059,  ..., -0.1020, -0.1020, -0.1059],
         [-0.1020, -0.1020, -0.1020,  ..., -0.0980, -0.0980, -0.1020],
         [-0.1020, -0.1020, -0.1020,  ..., -0.0980, -0.0980, -0.1020],
         ...,
         [-0.1020, -0.0980, -0.1020,  ..., -0.1020, -0.1020, -0.1059],
         [-0.1020, -0.0980, -0.1020,  ..., -0.1020, -0.1020, -0.1059],
         [-0.1020, -0.0980, -0.1020,  ..., -0.1020, -0.1020, -0.1059]],

        [[-0.1020, -0.1020, -0.1020,  ..., -0.0980, -0.0980, -0.1020],
         [-0.0980, -0.0980, -0.0980,  ..., -0.0941, -0.0941, -0.0980],
         [-0.0980, -0.0980, -0.0980,  ..., -0.0941, -0.0941, -0.0980],
         ...,
         [-0.0980, -0.0941, -0.0980,  ..., -0.0980, -0.0980, -0.1020],
         [-0.0980, -0.0941, -0.0980,  ..., -0.0980, -0.0980, -0.1020],
         [-0.0980, -0.0941, -0.0980,  ..., -0.0980, -0.0980, -0.1020]],

        [[-0.0941, -0.0941, -0.0941,  ..., -0.0902, -0.0902, -0.0941],
         [-0.0902, -0.0902, -0.0902,  ..., -0.0863, -0.0863, -0.0902],
         [-0.0902, -0.0902, -0.0902,  ..., -0.0863, -0.0863, -0.0902],
         ...,
         [-0.0902, -0.0863, -0.0902,  ..., -0.0902, -0.0902, -0.0941],
         [-0.0902, -0.0863, -0.0902,  ..., -0.0902, -0.0902, -0.0941],
         [-0.0902, -0.0863, -0.0902,  ..., -0.0902, -0.0902, -0.0941]]])

torchvision.transforms.Copmose(transforms)

#nice 是Image
com=transforms.Compose([
    transforms.ToTensor(),
    # transforms.ToPILImage()
])

print(com)
a=com(nice)
print(type(a))

Compose(
    ToTensor()
)
<class 'torch.Tensor'>

com=transforms.Compose([
    transforms.ToTensor(),
    transforms.ToPILImage()
])

print(com)
a=com(nice)
print(type(a))

Compose(
    ToTensor()
    ToPILImage()
)
<class 'PIL.Image.Image'>