mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 01:52:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Double learning rate of exp-scale units

Browse Source
This commit is contained in:
Daniel Povey 2022-03-04 00:42:37 +08:00
parent 5c177fc52b
commit 23b3aa233c
3 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
egs/librispeech/ASR/conformer_ctc/subsampling.py
View File

@ -48,12 +48,12 @@ class Conv2dSubsampling(nn.Module):
in_channels=1, out_channels=odim, kernel_size=3, stride=2 in_channels=1, out_channels=odim, kernel_size=3, stride=2
), ),
nn.ReLU(), nn.ReLU(),
ExpScale(odim, 1, 1, speed=2.0), ExpScale(odim, 1, 1, speed=4.0),
nn.Conv2d( nn.Conv2d(
in_channels=odim, out_channels=odim, kernel_size=3, stride=2 in_channels=odim, out_channels=odim, kernel_size=3, stride=2
), ),
nn.ReLU(), nn.ReLU(),
ExpScale(odim, 1, 1, speed=2.0), ExpScale(odim, 1, 1, speed=4.0),
) )
self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim) self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim)
self.out_norm = nn.LayerNorm(odim, elementwise_affine=False) self.out_norm = nn.LayerNorm(odim, elementwise_affine=False)

4
egs/librispeech/ASR/transducer_stateless/conformer.py
View File

@ -157,7 +157,7 @@ class ConformerEncoderLayer(nn.Module):
self.feed_forward = nn.Sequential( self.feed_forward = nn.Sequential(
nn.Linear(d_model, dim_feedforward), nn.Linear(d_model, dim_feedforward),
Swish(), Swish(),
ExpScale(dim_feedforward, speed=2.0), ExpScale(dim_feedforward, speed=4.0),
nn.Dropout(dropout), nn.Dropout(dropout),
nn.Linear(dim_feedforward, d_model), nn.Linear(dim_feedforward, d_model),
) )
@ -165,7 +165,7 @@ class ConformerEncoderLayer(nn.Module):
self.feed_forward_macaron = nn.Sequential( self.feed_forward_macaron = nn.Sequential(
nn.Linear(d_model, dim_feedforward), nn.Linear(d_model, dim_feedforward),
Swish(), Swish(),
ExpScale(dim_feedforward, speed=2.0), ExpScale(dim_feedforward, speed=4.0),
nn.Dropout(dropout), nn.Dropout(dropout),
nn.Linear(dim_feedforward, d_model), nn.Linear(dim_feedforward, d_model),
) )

2
egs/librispeech/ASR/transducer_stateless/train.py
View File

@ -110,7 +110,7 @@ def get_parser():
parser.add_argument( parser.add_argument(
"--exp-dir", "--exp-dir",
type=str, type=str,
default="transducer_stateless/specaugmod_baseline_randcombine1_expscale", default="transducer_stateless/specaugmod_baseline_randcombine1_expscale2",
help="""The experiment dir. help="""The experiment dir.
It specifies the directory where all training related It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved files, e.g., checkpoints, log, etc, are saved