Start ff modules with larger initial_scale

This commit is contained in:
Daniel Povey 2022-12-29 18:50:12 +08:00
parent fbdb12cf77
commit c6bad1ee4f

View File

@ -1405,7 +1405,7 @@ class FeedforwardModule(nn.Module):
self.activation = SwooshL()
self.dropout = Dropout2(dropout)
self.out_proj = ScaledLinear(feedforward_dim, embed_dim,
initial_scale=0.01)
initial_scale=0.1)
self.out_whiten = Whiten(num_groups=1,
whitening_limit=_whitening_schedule(7.5),