Anomaly with RandomGrayScale tests about torchgeo HOT 2 OPEN

In theory, the following change should allow the test to pass but it does not.

Actually, and fully realizing that I was the one who wrote these tests, I see no reason why these tests should pass. As an example, imagine an image with 3 channels. The first channel is white (255). All other channels are black (0). If we use weights=[0, 0, 1] or weights=[1, 0, 0], we will see completely different outputs (all black in the first case, same image in the second case). I actually think we should remove the test for having the same sum at the end.

from torchgeo.

Passing sample through our AugmentationSequential is changing the value of sample["image"] to be equal to that of output["image"].

This looks like a bug in our AugmentationSequential wrapper. The upstream Kornia implementation does not suffer from this bug. This can be tested with:

import kornia.augmentation as K
import torch
import torchgeo.transforms as T

image = torch.randn(1, 1, 3, 3)
aug_dict = K.AugmentationSequential(K.RandomHorizontalFlip(p=1.0), data_keys=None)
data = {"image": image}

print("\nKornia:")
print(data["image"])
out = aug_dict(data)
print(data["image"])
print(out["image"])

image = torch.randn(1, 1, 3, 3)
aug_dict = T.AugmentationSequential(K.RandomHorizontalFlip(p=1.0), data_keys=["image"])
data = {"image": image}

print("\nTorchGeo:")
print(data["image"])
out = aug_dict(data)
print(data["image"])
print(out["image"])

I wouldn't devote too much energy to fixing this. We should instead remove our implementation.

from torchgeo.