Use the same schedule for in_proj as out_proj. Only affects a couple of modules.
This commit is contained in:
parent
9e7add6be8
commit
9752778ee6
@ -347,7 +347,8 @@ def _aux_grad_scale() -> float:
|
|||||||
def _aux_grad_prob_out() -> ScheduledFloat:
|
def _aux_grad_prob_out() -> ScheduledFloat:
|
||||||
return ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125))
|
return ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125))
|
||||||
def _aux_grad_prob_in() -> ScheduledFloat:
|
def _aux_grad_prob_in() -> ScheduledFloat:
|
||||||
return ScheduledFloat((0.0, 0.25), (1000.0, 0.0))
|
return ScheduledFloat((0.0, 0.25), (1000.0, 0.05), (8000.0, 0.0125))
|
||||||
|
#return ScheduledFloat((0.0, 0.25), (1000.0, 0.0))
|
||||||
|
|
||||||
|
|
||||||
class ZipformerEncoderLayer(nn.Module):
|
class ZipformerEncoderLayer(nn.Module):
|
||||||
|
|||||||
Reference in New Issue
Block a user