diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py index 140638d3c..b406da0c7 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py @@ -1306,7 +1306,7 @@ class AttentionSqueeze(nn.Module): max_factor=0.02, min_prob=0.1, ) - self.bottleneck_activation = DoubleSwish() # in bottleneck + self.bottleneck_activation = nn.Tanh() # in bottleneck self.activation = Identity() # for diagnostics # the next two balancers are only to stop parameter-magnitude 'drift': we have