The cropping may have some issues

[Shiyu Xuan]

ch0 = max(int(height * h0) - h//2, 0)
ch1 = min(int(height * h1) - h//2, height - h)
In your implementation, ch0 can be larger than ch1. For example, ch0 and ch1 are both near 1.0.
I think this situation may cause some errors.

[xyupeng]

Hi @SY-Xuan,
thank you for pointing out the problem. I will try to restrict int(height * h0) - h//2 and int(height * h1) - h//2 within [0, height - h] and rerun some main experiments. And did you catch such errors in your training?

[Shiyu Xuan]

Yeah! If you use the uniform sampling like
'i = torch.randint(ch0, ch1 + 1, size=(1,)).item()'
There can be some issues.
ch0 = int(height * h0) - h//2 can be larger than ch1 = height - h.
I think you did not catch the issue because you use the Beta distribution to sample here.
Even though the ch0 > ch1, it will not have an issue.
This situation may produce some unknown behaviors.

[xyupeng]

Already fixed. If any further question, feel free to reopen the issue.

[Shiyu Xuan]

Thanks for your kindly replying. One more thing, I wonder if there is any improvement in the performance after fixing this bug?