there may be a small bug when the quality is 100

[brightendavid]

thank you for your work.
when i train my net in watermark,it gets NaN when the quality setted to be 100.However, it is just a small bug,and not affect the whole.

[jj199603]

I also find this bug. This is because the quality factor will be 0 if you set quality too 100. Then, the factor will be an DIVISION BY ZERO error in quantization. My solution is add 0.0001 to the factor:
quality/100. + 0.0001 in function quality_to_factor.
Of course, if you use a random quality, such as torch.rand(1)*100, this bug is no problem, because the range is [0, 100). In my experiment, I set this value = 100. - torch.rand(1)*50, although the range is (50, 100], since the probability is continuous, the probability of getting to 100 is almost zero.

[jj199603]

This is because the quality factor will be 0 if you set quality too 100. Then, the factor will be an DIVISION BY ZERO error in quantization. My solution is add 0.0001 to the factor:

I also find this bug. This is because the quality factor will be 0 if you set quality too 100. Then, the factor will be an DIVISION BY ZERO error in quantization. My solution is add 0.0001 to the factor:
quality/100. + 0.0001 in function quality_to_factor.
Of course, if you use a random quality, such as torch.rand(1)*100, this bug is no problem, because the range is [0, 100). In my experiment, I set this value = 100. - torch.rand(1)*50, although the range is (50, 100], since the probability is continuous, the probability of getting to 100 is almost zero.