From 7de7753ea2421d415b64d13e6dd26f494a33ee76 Mon Sep 17 00:00:00 2001 From: Daniel Povey Date: Mon, 12 Dec 2022 15:58:36 +0800 Subject: [PATCH] Change DoubleSwish to SwooshR in Conv2dSubsampling, double max_abs limits. --- .../ASR/pruned_transducer_stateless7/zipformer.py | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py index 65c2ffb0d..aff5a7704 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/zipformer.py @@ -1775,11 +1775,11 @@ class Conv2dSubsampling(nn.Module): kernel_size=3, padding=(0, 1), # (time, freq) ), - ScaleGrad(0.1), + ScaleGrad(0.2), ActivationBalancer(layer1_channels, channel_dim=1, - max_abs=0.5), - DoubleSwish(), + max_abs=1.0), + SwooshR(), nn.Conv2d( in_channels=layer1_channels, out_channels=layer2_channels, @@ -1789,8 +1789,8 @@ class Conv2dSubsampling(nn.Module): ), ActivationBalancer(layer2_channels, channel_dim=1, - max_abs=2.0), - DoubleSwish(), + max_abs=4.0), + SwooshR(), nn.Conv2d( in_channels=layer2_channels, out_channels=layer3_channels, @@ -1799,8 +1799,8 @@ class Conv2dSubsampling(nn.Module): ), ActivationBalancer(layer3_channels, channel_dim=1, - max_abs=2.0), - DoubleSwish(), + max_abs=4.0), + SwooshR(), ) out_height = (((in_channels - 1) // 2) - 1) // 2