Make nonlin_skip_rate nonzero and end after 20k iters; remove peak at 8k iteras of NonlinAttentionModule balancer2 min_abs.

This commit is contained in:
Daniel Povey 2022-12-13 19:17:38 +08:00
parent 22204450db
commit 117d418e27

View File

@ -399,7 +399,7 @@ class ZipformerEncoderLayer(nn.Module):
# to work correctly.
layer_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.05), default=0),
dynamic_skip_rate: FloatLike = ScheduledFloat((0.0, 0.2), (4000.0, 0.0), default=0),
nonlin_skip_rate: FloatLike = 0.0,
nonlin_skip_rate: FloatLike = ScheduledFloat((0.0, 0.1), (20000, 0.0), default=0),
const_attention_rate: FloatLike = ScheduledFloat((0.0, 0.25), (4000.0, 0.025), default=0),
bypass_min: FloatLike = ScheduledFloat((0.0, 0.75), (20000.0, 0.2), default=0),
bypass_max: FloatLike = 1.0,