From e3aed2e219770332822c853e6bc62555c6ff4a2d Mon Sep 17 00:00:00 2001
From: dohe0342 <kimdohe1070@gmail.com>
Date: Mon, 9 Jan 2023 19:51:14 +0900
Subject: [PATCH] from local

---
 .../ASR/incremental_transf/.conformer.py.swp  | Bin 114688 -> 114688 bytes
 .../ASR/incremental_transf/.model.py.swp      | Bin 24576 -> 24576 bytes
 .../ASR/incremental_transf/model.py           |   9 ++++++++-
 3 files changed, 8 insertions(+), 1 deletion(-)

diff --git a/egs/librispeech/ASR/incremental_transf/.conformer.py.swp b/egs/librispeech/ASR/incremental_transf/.conformer.py.swp
index e89a7c65acbb2d1859fe913538e87063dc16f2cc..716a29e386cc4f02ea81f4770810f561421157a8 100644
GIT binary patch
delta 37
rcmZo@U~gz(7f&(>^Ym4))H7fJ0s#hw6CZXbGcfQrif<KXe7PS0wS5X(

delta 37
rcmZo@U~gz(7f&(>^Ym4))H7fJ0s#hwLmzf0Z~f2HD85yk@#TI1#VQO4

diff --git a/egs/librispeech/ASR/incremental_transf/.model.py.swp b/egs/librispeech/ASR/incremental_transf/.model.py.swp
index 0342b02ceb13045ab2eae006bbcda8cb51ab6750..93db73defa70a915f53fc79762cc56dc517f7f5a 100644
GIT binary patch
delta 402
zcmZoTz}Rqral?CNMuyEFm>YN)Wi}TIOxEZ9&Begb&dtEUAi%&NGg;6<I&wK5149ZB
z2LrJe5Q_rwK3)cfnLykL#4bRr3dC$c{Dg;rVLK4_193VKy8y8~5Pt%icmaqH0r4gv
z-U!6)n_C_Jb4^zCv|?6QSKqwQ=K^EB0uTT>3?LdA<QAvK=j0a`E7&R&r{<*T0r?sX
z$O_W(ixe{RN>YpB^Giz#N=tMU3QC|%ptwR+W`PDoIw!FbuBaHS3Q2^eq*&8x^28vs
zD45xBOQ5zJzy+`v2Xh@nDm5=TKP9zD1LzPWD@rB{c!^H_;3qwKjlIxh2Y<E64aV%7
Jh5g?O002hIZu$TK

delta 219
zcmZoTz}Rqral?CN#?Z|lm>YN)#WoiTOxEXR=3!tc2Vw>R1_rUof)3J=^Z6JU<^gdg
z5UT(&D-b{9Wnef0#M6N|ABf$7SOtjx19e^k;tfE&2#6N~G03z^Ag%ynHz1Y-;&0px
z3@10YI{fFFEa1s9Inh;V@(maH$z2|blMUQNCZBeb28x+-6s4Aw7Ud~u06FrL54xF6
sUf?CV`GNOF#>umS%qDL%5}W+NPkizkd!fk={%VsOjM+B}`@a<c0A_AFCIA2c

diff --git a/egs/librispeech/ASR/incremental_transf/model.py b/egs/librispeech/ASR/incremental_transf/model.py
index 396e99408..29f520300 100644
--- a/egs/librispeech/ASR/incremental_transf/model.py
+++ b/egs/librispeech/ASR/incremental_transf/model.py
@@ -244,13 +244,19 @@ class Interformer(nn.Module):
         x: torch.Tensor,
         x_lens: torch.Tensor,
     ):
-        encoder_out, x_lens, layer_outputs = self.pt_encoder(x, 
+        encoder_out, x_lens, pt_layer_outputs = self.pt_encoder(x, 
                                                              x_lens, 
                                                              warmup=warmup, 
                                                              get_layer_output=True
                                             )
 
+        inter_layer_outputs = self.inter_encoder(pt_layer_outputs)
+        loss = 0
+        
         with torch.cuda.amp.autocast(enabled=False):
+            for inter_output, pt_output in zip(inter_layer_outputs, pt_layer_outputs):
+                mse_loss = self.mse(
+            '''
             simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
                 lm=lm.float(),
                 am=am.float(),
@@ -263,5 +269,6 @@ class Interformer(nn.Module):
                 delay_penalty=delay_penalty,
                 return_grad=True,
             )
+            '''
 
         return (simple_loss, pruned_loss)