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

Let ratio of values to sigmoids be 8, not 2

Browse Source
This commit is contained in:
Daniel Povey 2022-11-28 21:50:11 +08:00
parent 643c547eec
commit 7018c722b5
1 changed files with 5 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py
View File

@ -1459,18 +1459,19 @@ class NonlinAttentionModule(nn.Module):
""" """
def __init__( def __init__(
self, channels: int, self, channels: int, ratio: int = 8,
) -> None: ) -> None:
super().__init__() super().__init__()
self.in_proj = nn.Linear(channels, channels + channels // 2, bias=True) assert channels % ratio == 0
self.in_proj = nn.Linear(channels, channels + channels // ratio, bias=True)
# balancer that goes before the sigmoid. Have quite a large min_abs value, at 2.0, # balancer that goes before the sigmoid. Have quite a large min_abs value, at 2.0,
# because we noticed that well-trained instances of this module have abs-value before the sigmoid # because we noticed that well-trained instances of this module have abs-value before the sigmoid
# starting from about 3, and poorly-trained instances of the module have smaller abs values # starting from about 3, and poorly-trained instances of the module have smaller abs values
# before the sigmoid. # before the sigmoid.
self.balancer = ActivationBalancer( self.balancer = ActivationBalancer(
channels // 2, channel_dim=-1, channels // ratio, channel_dim=-1,
min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)), min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)),
max_positive=1.0, max_positive=1.0,
min_abs=2.0, min_abs=2.0,
@ -1512,7 +1513,7 @@ attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len)
s = self.balancer(s) s = self.balancer(s)
s = self.sigmoid(s) s = self.sigmoid(s)
s = s.unsqueeze(-1).expand(-1, -1, -1, 2).reshape(seq_len, batch_size, num_channels) s = s.unsqueeze(-1).expand(-1, -1, -1, ratio).reshape(seq_len, batch_size, num_channels)
x = x * s x = x * s