mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reduce kernel size of convnext2 from 7 to 5.

Browse Source
This commit is contained in:
Daniel Povey 2022-12-31 17:10:31 +08:00
parent c533c30442
commit d48b2ccb45
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
View File

@ -1704,10 +1704,10 @@ class ConvNeXt(nn.Module):
def __init__(self, def __init__(self,
channels: int, channels: int,
hidden_ratio: int = 3, hidden_ratio: int = 3,
kernel_size: Tuple[int, int] = (7, 7),
layerdrop_rate: FloatLike = None): layerdrop_rate: FloatLike = None):
super().__init__() super().__init__()
kernel_size = 7 padding = ((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2)
pad = (kernel_size - 1) // 2
hidden_channels = channels * hidden_ratio hidden_channels = channels * hidden_ratio
if layerdrop_rate is None: if layerdrop_rate is None:
layerdrop_rate = ScheduledFloat((0.0, 0.2), (20000.0, 0.015)) layerdrop_rate = ScheduledFloat((0.0, 0.2), (20000.0, 0.015))
@ -1717,8 +1717,8 @@ class ConvNeXt(nn.Module):
in_channels=channels, in_channels=channels,
out_channels=channels, out_channels=channels,
groups=channels, groups=channels,
kernel_size=7, kernel_size=kernel_size,
padding=(3, 3)) padding=padding)
self.pointwise_conv1 = nn.Conv2d( self.pointwise_conv1 = nn.Conv2d(
in_channels=channels, in_channels=channels,
@ -1869,9 +1869,9 @@ class Conv2dSubsampling(nn.Module):
SwooshR(), SwooshR(),
) )
self.convnext2 = nn.Sequential(ConvNeXt(layer3_channels), self.convnext2 = nn.Sequential(ConvNeXt(layer3_channels, kernel_size=(5, 5)),
ConvNeXt(layer3_channels), ConvNeXt(layer3_channels, kernel_size=(5, 5)),
ConvNeXt(layer3_channels)) ConvNeXt(layer3_channels, kernel_size=(5, 5)))
out_width = (((in_channels - 1) // 2) - 1) // 2 out_width = (((in_channels - 1) // 2) - 1) // 2