PIL: Image.fromarray(img.astype('uint8'), mode='RGB') 返回灰度图像

Question

我已将大小为 torch.Size([3, 28, 28]) 的 pytorch 张量转换为大小为 (28, 28, 3) 的 numpy 数组，这似乎没有任何问题。然后我尝试使用img = Image.fromarray(img.astype('uint8'), mode='RGB') 将其转换为PIL 图像，但返回的img 的尺寸是(28, 28)，而我期望它是(28, 28, 3)（或(3, 28, 28)）。我不明白为什么会这样。正如其他海报在线建议的那样，我确保转换为 uint8 并使用 RGB 模式，但是这些都没有帮助（也没有使用 np.ascontiguousarray）。

PIL 版本 1.1.7

# This code implements the __getitem__ function for a child class of datasets.MNIST in pytorch
# https://pytorch.org/docs/stable/_modules/torchvision/datasets/mnist.html#MNIST

img, label = self.data[index], self.targets[index]
assert img.shape == (3, 28, 28), \
       (f'[Before PIL] Incorrect image shape: expecting (3, 28, 28),'
        f'received {img.shape}')

print('Before reshape:', img.shape)    # torch.Size([3, 28, 28])
img = img.numpy().reshape(3, 28, 28)
img = np.stack([img[0,:,:], img[1,:,:], img[2,:,:]], axis=2)
print('After reshape:', img.shape)     # (28, 28, 3)

# doing this so that it is consistent with all other datasets
# to return a PIL Image
img = Image.fromarray(img.astype('uint8'), mode='RGB') # Returns 28 x 28 image

assert img.size == (3, 28, 28), \
       (f'[Before Transform] Incorrect image shape: expecting (3, 28, 28), '
        f'received {img.size}')

编辑：这是一个最小的例子。如果有任何帮助，我将把上面的内容留作上下文。

from PIL import Image
import numpy as np

img = np.random.randn(28, 28, 3)
img = Image.fromarray(img.astype('uint8'), mode='RGB') # Returns 28 x 28 image

assert img.size == (28, 28, 3), \
       (f'[Before Transform] Incorrect image shape: expecting (3, 28, 28), '
        f'received {img.size}')

AssertionError: [Before Transform] Incorrect image shape: expecting (3, 28, 28), received (28, 28)

Answer 1

我想你想要这个，其中 RGB 值的范围是 0..255 范围内的整数：

import numpy as np
from PIL import Image

# Make random 28x28 RGB image
img =np.random.randint(0,256,(28,28,3), dtype=np.uint8)

# Convert to PIL Image
pImg=Image.fromarray(img, mode='RGB')

现在检查我们有什么：

In [19]: pImg                                                                                       
Out[19]: <PIL.Image.Image image mode=RGB size=28x28 at 0x120CE9CF8>

然后保存：

pImg.save('result.png')