Reduce initial clamp_min for bypass_scale from 1.0 to 0.5.

This commit is contained in:
Daniel Povey 2022-10-30 14:43:02 +08:00
parent e9c69d8477
commit e4a22bbe96

View File

@ -321,7 +321,7 @@ class ZipformerEncoderLayer(nn.Module):
return self.bypass_scale
# hardcode warmup period for bypass scale
warmup_period = 4000.0
initial_clamp_min = 1.0
initial_clamp_min = 0.5
final_clamp_min = 0.2
if self.batch_count > warmup_period:
clamp_min = final_clamp_min
@ -330,8 +330,6 @@ class ZipformerEncoderLayer(nn.Module):
(self.batch_count / warmup_period) * (initial_clamp_min - final_clamp_min))
return self.bypass_scale.clamp(min=clamp_min, max=1.0)
def forward(
self,
src: Tensor,