torchvision.transforms.functional.to_pil_image

[liangsizhuang]

The function, torchvision.transforms.functional.to_pil_image, transforms FloatTensor to uint8 internally. Hence, it is impossible to use mode = 'F'. This can be frustrating when dealing with high bit-depth images.

I was working with spectrograms and the spectrograms had high resolution (more than 8 bits per channel).

Problem Demonstration 1

Here is a demonstration of what was happening.

import torch
import torchvision.transforms
X1 = torch.Tensor(10, 10)
X1.normal_(0.5, 0.5)
print(X1.dtype)
X2 = torchvision.transforms.functional.to_pil_image(X1, mode = 'F') # This is going to be error.
Output: ValueError: Incorrect mode (F) supplied for input type <class 'numpy.dtype'>. Should be L

Problem Demonstration 2

If we use mode L, we have

X2 = torchvision.transforms.functional.to_pil_image(X1)
X3 = torchvision.transforms.functional.to_tensor(X2)
print((X1 - X3).abs().max())
Output: tensor(2.0097)

This means that X3 is significantly different from X1.

Expected Behavior

This is what it should be

from PIL import Image
X4 = Image.fromarray(X1.numpy(), mode = 'F')
X5 = torchvision.transforms.functional.to_tensor(X4)
print((X1 - X5).abs().max())
Output: tensor(0.)