From 121f7e2a451e2cc7c2958dce483e08c379081d1c Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Mon, 28 Nov 2022 12:10:08 +0800 Subject: [PATCH] Documentation fix. --- .../ASR/pruned_transducer_stateless7/zipformer.py | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py index ef75b892a..c771b2895 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py @@ -1314,11 +1314,11 @@ class AttentionSqueeze(nn.Module): self.bottleneck_activation = TanSwish() # in bottleneck self.activation = Identity() # for diagnostics - # the next two balancers are only to stop parameter-magnitude 'drift': we have - # too many degrees of freedom for the scales of the various activations. - # Make them run with very low probability, since only a small application of - # these balancers should be enough to stop such "drift"; and, for speed, - # put no limitation on the signs (so: min_positive=0, max_positive=1). + # the reason for the min_abs and max_abs limits on the next two + # balancers are only to stop parameter-magnitude 'drift': we have too + # many degrees of freedom for the scales of the various activations. + # Make them run with very low probability, since only a small + # application of these balancers should be enough to stop such "drift". self.scale_balancer = ActivationBalancer( embed_dim, channel_dim=-1, min_positive=0.2, max_positive=0.8,