search

snap-stanford / orca

[ICLR 2022] Open-World Semi-Supervised Learning
92 stars 12 forks source link

MarginLoss different from the paper #12

Closed huangyjhust closed 2 years ago

huangyjhust commented 2 years ago

Hi greatly appreciate your code and your research work! I tried your code and found it significantly improved open-set accuracy on multiple datasets, but I am confused by the difference between the MarginLoss in the codebase and the MarginLoss in the paper. The code "x_m = x - self.m self.s" where self.m = 0.2 and self.s = 10, accelerates the learning process of seen classes instead slowing down seen class learning and waiting for unseen class clustering during the begining epochs. From my understanding it should be "x_m = x + self.m self.s". Am I right about this point?

huangyjhust commented 2 years ago

ooooh great found the minus here: m = min(m, 0.5) ce = MarginLoss(m=-1*m)