Further tune lr scales; increase base-lr

This commit is contained in:
Daniel Povey 2023-01-06 13:34:48 +08:00
parent f6f088489d
commit 5564a0efb0

View File

@ -209,7 +209,7 @@ class Zipformer(EncoderInterface):
)
# we are adding a new attribute here.
# this will be interpreted by get_named_parameter_groups_with_lrs().
encoder.lr_scale = downsampling_factor[i] ** -0.25
encoder.lr_scale = downsampling_factor[i] ** -0.2
encoders.append(encoder)
self.encoders = nn.ModuleList(encoders)
@ -450,7 +450,7 @@ class ZipformerEncoderLayer(nn.Module):
dropout)
self.nonlin_attention = NonlinAttention(embed_dim,
hidden_channels=embed_dim // 4)
hidden_channels=embed_dim // 4)
self.conv_module = ConvolutionModule(embed_dim,
@ -1480,7 +1480,7 @@ class NonlinAttention(nn.Module):
) -> None:
super().__init__()
self.lr_scale = 0.9
self.lr_scale = 0.95
self.hidden_channels = hidden_channels