Asymmetric Loss

[tibrinkmann]

ASL implementation and config file. Default values from original paper/implementation.
Best parameters in experiments: gamma_pos = gamma_neg = 0 and clip = 0.15

[sfluegel05]

Thanks for the PR. I adjusted the torch grad handling. In your implementation (and the original implementation), the gradient gets disabled and enabled globally. This means that if someone has disabled grad calculation outside of the ASL function, that will be overwritten. With the context, this doesn't happen.

Also, for some reason, the single-label variant doesn't use the disable-grad flag. I don't see why. @tibrinkmann did you test the ASL with and / or without the disabled gradient? That seems to be quite an important difference. Also, is there any reason not to use the optimized version?