Would it be possible to calculate b from tau_threshold?

Hi,

I'm trying to adapt your method to my deep learning problem. I noticed that you set the b as hard parameter. But I wondered would it actually be possible to solve b from the inequality tau > (B+3b)/(3b) and use that? I actually tried this, but it seems to perform worse than this hard setting. Do you have any idea why this might fail?

best, Juuso