Change power of lr-schedule from -0.5 to -0.333

This commit is contained in:
Daniel Povey 2022-03-17 13:18:58 +08:00
parent 953aecf5e3
commit 8a8134b9e5

View File

@ -391,7 +391,8 @@ class Noam(object):
return ( return (
self.factor self.factor
* self.model_size ** (-0.5) * self.model_size ** (-0.5)
* min(step ** (-0.5), step * self.warmup ** (-1.5)) * self.warmup ** (-0.5 - -0.333)
* min(step ** (-0.333), step * self.warmup ** (-1.333))
) )
def zero_grad(self): def zero_grad(self):