Tensor, numpy, PIL format conversion and image display

a key

cv.imshow() ： Display image is BGR Format
plt.imshow() ： Image display yes RGB Format

Tensor ： The stored data is distributed in [0, 1]
Numpy ： The stored data is distributed in [0, 255]

CIFAR-10 Data sets

The data set is RGB Format ;
In the use of opencv It needs to be converted into BGR Format ;
In the use of plt Display time There is no need to Transformation format ;

Example ：

dict = unpickle('./dataset/cifar-10-batches-py/test_batch')

img = dict[b'data']
image = img[0]
image = image.reshape(3, 32, 32).transpose(1, 2, 0)
cv_show('image', image)

r, g, b = cv.split(image)
pic = cv.merge([b, g, r])
cv_show('pic', pic)

The picture on the left is the original , The picture on the right is a distorted picture ：

format conversion

import torch
import numpy as np

# convert numpy to tensor or vise versa
np_data = np.arange(6).reshape((2, 3))
torch_data = torch.from_numpy(np_data)  # ndarray Convert to Tensor
tensor2array = torch_data.numpy()  # Tensor Convert to ndarray

test = out.cpu().detach().numpy() #  from CUDA Upper Tensor Convert to ndarray

Tensor ==> Numpy

import torch
import torchvision
import pickle
import cv2 as cv

transform_tensor = torchvision.transforms.Compose([
    torchvision.transforms.ToTensor(),
])

transform_picture = torchvision.transforms.Compose([
    torchvision.transforms.ToTensor(),                   #  convert to Tensor Format 
    torchvision.transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])


def cv_show(name, img):
    cv.imshow(name, img)
    cv.waitKey(0)
    cv.destroyAllWindows()


def unpickle(file):
    with open(file, 'rb') as fo:
        dict = pickle.load(fo, encoding='bytes')
    return dict


dict = unpickle('./dataset/cifar-10-batches-py/test_batch')

img = dict[b'data']
image = img[0]
image = image.reshape(3, 32, 32).transpose(1, 2, 0)
print(image)

image_Tensor = transform_tensor(image).unsqueeze(0)
print(image_Tensor)            #  There is no data normalization operation 

image_Tensor_Nor = transform_picture(image).unsqueeze(0)
print(image_Tensor_Nor)        #  There is data normalization operation 

[[[158 112  49]
  [159 111  47]
  [165 116  51]
  ...
  [ 24  77 124]
  [ 34  84 129]
  [ 21  67 110]]]
  
tensor([[[[0.6196, 0.6235, 0.6471,  ..., 0.5373, 0.4941, 0.4549],
          [0.5961, 0.5922, 0.6235,  ..., 0.5333, 0.4902, 0.4667],
          [0.5922, 0.5922, 0.6196,  ..., 0.5451, 0.5098, 0.4706],
          ...,
          [0.6941, 0.5804, 0.5373,  ..., 0.5725, 0.4235, 0.4980],
          [0.6588, 0.5804, 0.5176,  ..., 0.5098, 0.4941, 0.4196],
          [0.6275, 0.5843, 0.5176,  ..., 0.4863, 0.5059, 0.4314]]]])
          
tensor([[[[ 0.6338,  0.6531,  0.7694,  ...,  0.2267,  0.0134, -0.1804],
          [ 0.5174,  0.4981,  0.6531,  ...,  0.2073, -0.0060, -0.1223],
          [ 0.4981,  0.4981,  0.6338,  ...,  0.2654,  0.0910, -0.1029],
          ...,
          [ 1.2319,  0.6661,  0.4515,  ...,  0.6271, -0.1143,  0.2564],
          [ 1.0563,  0.6661,  0.3540,  ...,  0.3149,  0.2369, -0.1338],
          [ 0.9003,  0.6856,  0.3540,  ...,  0.1979,  0.2954, -0.0753]]]])

Tensor turn Numpy ： （ For displaying images ）

mean = (0.4914, 0.4822, 0.4465)
std = (0.2023, 0.1994, 0.2010)

def tensor_numpy(image):
    clean = image.clone().detach().cpu().squeeze(0)        #  Get rid of batch passageway  (batch, C, H, W) --> (C, H, W)
    clean[0] = clean[0] * std[0] + mean[0]                 #  Data De normalization 
    clean[1] = clean[1] * std[1] + mean[1]
    clean[2] = clean[2].mul(std[2]) + mean[2]
    clean = np.around(clean.mul(255))                     #  Convert to color 255 [0, 1] --> [0, 255]
    clean = np.uint8(clean).transpose(1, 2, 0)            #  Change to three channels  (C, H, W) --> (H, W, C)
    r, g, b = cv.split(clean)                             # RGB  Channel switching 
    clean = cv.merge([b, g, r])
    return clean

Num = tensor_numpy(image_Tensor_Nor)
print(Num)

If you use cv.imshow() You need to use the RGB Channel switching ;
If you use plt.imshow() There is no need to use the above RGB Channel switching ;

Example ：

plt.imshow('image', image)
plt.show()

RGB Format ：

[[[158 112  49]
  [159 111  47]
  [165 116  51]
  ...
  [ 24  77 124]
  [ 34  84 129]
  [ 21  67 110]]]

RGB Channel switching ：

r, g, b = cv.split(image)
pic = cv.merge([b, g, r])
plt.imshow('image', pic)
plt.show()

BGR Format ：

[[[ 49 112 158]
  [ 47 111 159]
  [ 51 116 165]
  ...
  [124  77  24]
  [129  84  34]
  [110  67  21]]]

PyTorch save_image()

torchvision.utils.save_image()

Finally, let's take a look at pytorch How does the built-in function perform format conversion and save pictures ：

def save_image(
    tensor: Union[torch.Tensor, List[torch.Tensor]],
    fp: Union[Text, pathlib.Path, BinaryIO],
    format: Optional[str] = None,
    **kwargs
) 

    grid = make_grid(tensor, **kwargs)
    # Add 0.5 after unnormalizing to [0, 255] to round to nearest integer
    ndarr = grid.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy()
    im = Image.fromarray(ndarr)
    im.save(fp, format=format)