From 87b8f78d0752e873ce2d1ae95bd55193eb390784 Mon Sep 17 00:00:00 2001
From: dohe0342 <kimdohe1070@gmail.com>
Date: Sun, 8 Jan 2023 18:53:43 +0900
Subject: [PATCH] from local

---
 .../.conformer.py.swp                         | Bin 94208 -> 94208 bytes
 .../conformer.py                              |  10 ++++++++--
 2 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/.conformer.py.swp b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/.conformer.py.swp
index a2e17c052797d2c3d5ea8570e4c0e7873ce9634e..22d5cb27043c6d00af788c9a7c6b81933748e704 100644
GIT binary patch
delta 496
zcmYk$ze@sP7zgm@inBCL8v@JahJrgS&7p{(O(Ez{h>)vJJMg^get;FIZ8;Sj(bCYL
zP+U>q)>K<9=9W_+6gd_|_4O3#g%7;X%lmwvr_y9AP4+?yF=1yXWWTUKLqv7k+m4h=
zK_V5>kb(%jjuPF02?k6<D?sGHKCHk1F!1Onx`YylFa@m<qFboKA*3M%<3RB4Lm$+9
z6sL1s4nToT2t&h5^bA)ZLlitvcb?k5VY~797A#l_x6I{q)zUJ!pMJ~Cr1SB0#tx#g
zkAAH{6`QcwkU|JO8Dmx_9#GYw3JWj~9$>F6=Gb2*u6K$^Hl%zuo)>o|JyTG1g)b)D
zm+nAA-jf9C|L?RFxmd5skaLPEZ}DPY+ZLnU2-Pw*%Ut6nEEh`((SKRJ8J3coO|A?5
m%&e+&o!>K@o5RTkx0H0zlz6w@eHWJc9i6S`Zterj;>I6*oqNCl

delta 316
zcmXZWze@sf9LDkIOC9N@bxIJ~a!}CLT1%5cLw`geR0yS}xCqj1H3S7mxCH$R))7KY
zHMxlh1r6mA7&P42FR=GPA9%s%y*=K*@&;C`?3i?L;P|iB&xF+X<Kg^aV_GUAhb%I9
zo09HPMFoo(CZ%iCu?_>Fn~+*)U<(TvPD*VY;|MuqF@upUeV~h5Tg$rS@vw(HX7Lo4
z9&v>oEMW}&pcmRP|K;;NRV~!F-8FZ&QmU0VqiFsXOeP4wF0An={nJq`xzE5}G}MvO
j=P9!>7gYz(c)$rtSh=a0vqbZ|XzpU}|Ft_wvzq?{4yZ?a

diff --git a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/conformer.py
index 643858af0..984f23105 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/conformer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless_gtrans/conformer.py
@@ -86,7 +86,7 @@ class Conformer(EncoderInterface):
         short_chunk_size: int = 25,
         num_left_chunks: int = -1,
         causal: bool = False,
-        group_size: int = 12,
+        group_num: int = 12,
     ) -> None:
         super(Conformer, self).__init__()
 
@@ -125,7 +125,7 @@ class Conformer(EncoderInterface):
         self.encoder = ConformerEncoder(encoder_layer, num_encoder_layers)
         self._init_state: List[torch.Tensor] = [torch.empty(0)]
 
-        self.group_size = group_size
+        self.group_num = group_num
         self.group_layer_num = int(self.encoder_layers // self.group_size)
         self.alpha = nn.Parameter(torch.rand(self.group_size))
         self.sigmoid = nn.Sigmoid()
@@ -202,6 +202,12 @@ class Conformer(EncoderInterface):
 
         x = x.permute(1, 0, 2)  # (T, N, C) ->(N, T, C)
         layer_output = [x.permute(1, 0, 2) for x in layer_output]
+        
+        x = 0
+        for enum, alpha in enumerate(self.alpha):
+            x += self.sigmoid(alpha*layer_output[(enum+1)*self.group_layer_num-1])
+
+        x = self.layer_norm(x/self.group_size)
 
         return x, lengths