RandomErasing is buggy

[Zhaoyi-Yan]

@zhunzhong07 @fmassa
https://github.com/pytorch/vision/blob/master/torchvision/transforms/transforms.py#L1256-L1263

Here, you assume that img_h < img_w implictly, which is true for most images. However, for some images img_h > img_w.

if h < img_h and w < img_w:

This line of code equals to:

h/w <(ratio0~ratio1)/(scale0~scale1)

where the right falls in the range of (ratio0/scale1, ratio1/scale0), here is (0.9, 16.7). Then when h is more than 17 times of w, the loop runs forever.

If the user sets custom ratio and scale, this may happen more frequenently. At least, more checks are deserved to add. Or maybe give warnings if this probablity is high.

[zhunzhong07]

Thanks for your question. I think your inference is not correct.

h/w <(ratio0~ratio1)/(scale0~scale1) is not correct.

h = int(round(math.sqrt(erase_area * aspect_ratio)))
w = int(round(math.sqrt(erase_area / aspect_ratio)))

=>>
h/w = sqrt(erase_area * aspect_ratio / (erase_area / aspect_ratio)) = aspect_ratio

=>>
h/w is in the range of (ratio[0], ratio[1]) The default is (0.3, 1/0.3)

Thus. h/w is unrelated to the scale.

[Zhaoyi-Yan]

@zhunzhong07 Hi, please try img_h = 300, img_w = 5, then you know the loop will run all the time.

[zhunzhong07]

@Zhaoyi-Yan Thanks for your pointing out. In such strict image size (img_h/img_w=60), I think you should set ratio[1] to a large value. You could set ratio according to the size of your own image dataset.

[zhunzhong07]

@Zhaoyi-Yan I think it would be better if we change
while True: to for attempt in range(100): and return the original image if there are no sufficient h and w selected?

[Zhaoyi-Yan]

Give a warning if it satisfies if h < img_h and w < img_w: with low probablity(less than a threshold). Give an error and correponding instruction if if h < img_h and w < img_w: never happens.

[fmassa]

@zhunzhong07

I think it would be better if we change
while True: to for attempt in range(100): and return the original image if there are no sufficient h and w selected?

This sounds good to me, but I think we can try for a few (say 10) iterations only, to keep things reasonably fast.

[zhunzhong07]

@fmassa Thanks for your suggestion. I will modify the code and create a PR.

[fmassa]

Fixed via #1095