[Feature Request] Un-Normalize Image Tensor

[varagrawal]

Basically the inverse of transforms.Normalize as this will allow us to visualize tensors during training more easily.

[karandwivedi42]

#281

[fmassa]

I agree with @karandwivedi42 and the comment that he linked to.

If you want to reverse the normalization, all you need to do is to use a new normalization, with slight modifications:

mean = torch.tensor([1, 2, 3], dtype=torch.float32)
std = torch.tensor([2, 2, 2], dtype=torch.float32)

normalize = T.Normalize(mean.tolist(), std.tolist())

unnormalize = T.Normalize((-mean / std).tolist(), (1.0 / std).tolist())

EDIT: fixed thanks to @karandwivedi42 comment

[fmassa]

Thanks @karandwivedi42 , my bad!
I'll fix my comment!

[danieltudosiu]

Hey,

This does not seem to work for me at all. I am using the following code:

import numpy as np
import matplotlib.pyplot as plt
from albumentations.augmentations import Normalize

mean = [-0.485 / 0.229, -0.456 / 0.224, -0.406 / 0.225]
std = [1.0 / 0.229, 1.0 / 0.224, 1.0 / 0.225]
a = np.load("/path/to/your/image)
plt.imshow(a)
plt.show()

norm = Normalize(always_apply=True)
a = norm(image=a)["image"]
plt.imshow(a)
plt.show()

denorm = Normalize(mean=mean, std=std, always_apply=True)
a = denorm(image=a)["image"]
plt.imshow(a)
plt.show()

Could you please tell me what am I doing wrong :(

[danieltudosiu]

For future reference this is the correct answer:

mean = (0.485, 0.456, 0.406)
std = (0.229, 0.224, 0.225)
a = np.load(
    "/home/danieltudosiu/storage/datasets/green_roof_retrofit/debugging/train/img_rgb/roofs_image_0002.npy"
)
plt.imshow(a)
plt.show()

norm = Normalize(mean=mean, std=std, always_apply=True)
a = norm(image=a)["image"]
plt.imshow(a)
plt.show()

denorm = Normalize(
    mean=[-m / s for m, s in zip(mean, std)],
    std=[1.0 / s for s in std],
    always_apply=True,
    max_pixel_value=1.0
)
a = (denorm(image=a)["image"]*255).astype(np.uint8)
plt.imshow(a)
plt.show()

[varunagrawal]

Your solution is pretty much what @fmassa laid out. If you look at his unnormalize method, he is computing the negative mean divided by the standard deviation.

[tkupek]

I also had issues with the solution above. Seems like max_pixel_value=1.0 and separately multiply with 255 from @danieltudosiu does the trick. You can also combine with in a simple Compose pipeline:

Compose(
        [
            Normalize(
                mean=tuple(-m / s for m, s in zip(mean, std)),
                std=tuple(1.0 / s for s in std),
                max_pixel_value=1.0,
            ),
            FromFloat(max_value=255, dtype="uint8"),
        ]
    )