From ab16ce6424105056a8a138aa75297b7136cc3021 Mon Sep 17 00:00:00 2001
From: dohe0342 <kimdohe1070@gmail.com>
Date: Thu, 16 Feb 2023 17:40:25 +0900
Subject: [PATCH] from local

---
 .../ASR/conformer_ctc2/.train.py.swp          | Bin 57344 -> 61440 bytes
 egs/librispeech/ASR/conformer_ctc2/train.py   |  52 +++++++++++++++++-
 2 files changed, 50 insertions(+), 2 deletions(-)

diff --git a/egs/librispeech/ASR/conformer_ctc2/.train.py.swp b/egs/librispeech/ASR/conformer_ctc2/.train.py.swp
index b7ed26d19b408f2d24a24ca0d2ec800f6f3a7a91..3d9b44bb3aa5b18f0f14fde69da420e0760141db 100644
GIT binary patch
delta 1905
zcmbW2&u<$=6vxLuk~E}Aokn%3K<GgZwgqg~O^Q@25r~3N6CqU!(4q>hCcEQ!#d_A-
znRVkH*!}^%nG>jp1GK0T65tCbR3L2;Jyi(71?T|@2`EPp;=*^=j<abK4vh5i?#{el
zGw;pwg=@JBcXGuw^Och?(s9FN7zM^Y{OztiyZUkQ_XA^-EW%v2TP+;!QK68@G`iL4
z!w08gNsRgagFTz?)ut~`eM0$nCs!~xmc<V&HG0$6N}1@9QO1hk?g(Qy!3}U8%z+Vb
zJCAwrK6nqD2hV^SI1UbhJ2}R_2Uoyla1p!=R=^B62!4Uvx4=d40yqQ^O2lqv7+c22
zN$_Bn(O}HdtA0qT;fJ0>p%P>mbDnQ=Z_aU5%i|rZ$)#HtN?S^E?JDis%4ntZVJP*r
zhkA66ekY}Jf`{7mWsP2Hh(M65h*KcA@+E1Xl<<VD>8NR#6t5d}%EO|-f=C(J8O4ea
zE%K$;p-^iPa2eS`(K%N)$O!{%1Gd+MJ|~2vDlFN)bclna=HN>uwjEJNC{4dDEF4jw
zEy!4&FV`*DE!#CZ5$Z~f4dc)Occo>ATG0}D?plB|6vA#8#lZ?fmsZs@6Pa?QdwX-{
z?4C;f1@3up9@Ymci7&TKEgtZuGU_3g6HkLMQeJ7P`^Wr|sWK1NRVC3Yl}f!|znsc;
z|Gr+?PZ*d;7J84}`0+@@9mc*GW9;g_e(i46_NNCJYk@Ps08fGua0{0p0#z^$ex79P
z8}Jd>0IT46@EjNgxArskHFyKegK_Zf1m?j96U=5;@X`TKgJWO;jDqhTVC)@m3A_n3
zI0Z_8z@Ouc{Qy1#mp}kcgNML0D1e`^sjK^XnfzB{y}v%sWhYW6)y-@gXCB{~GDoy6
zU!&yG2jAwN>+F``SlUjq8)v^<>DNRhKI884=JF0_Z*M-8p&nA7DG%$#sJk92wLGXc
zSN1)j%q&~mD3t3n(O$y$m*&8I>^%vA#cx|<k(zwN@<ZJUwdKnAH@OmQOCWwOk_?-c
z_ForL(Hzx1pGT;8qb|4(1EGv~xsDPPgVNekZsG#`hwVT>T~e71<#&w**T*!f+d{F?
zh5I5-q)gL1HWW++jcf{8?*7w$II;~DwnWf&6)wIyYrbY^v7sydI}rymo8E0-QMC*q
zZQl{t^lsWu5O!(lPD?Dga;L3fLpv>qD;4IOf^=7|79;ewg-4TMQ=R3WT<_9W28UKf
z3tU_x+ZgsRe9gy=F)@jKU?ZqlfDAQ!&q>X5NnzA)hPK%5g6Q%(aMzL!Om#n?jvof(
YaE%tLxBG9((Es|s=)|Vhb3(I!0nP0Wpa1{>

delta 253
zcmXBOy9<F~6bA6~exJ+W_I?Z5-YiPOH?m@oECw3`i;~UG;t$ZbSd>8qi_7~5ATeW>
z*yNJUWsyft{p$4eoH~tz+L)>CVkDQ!h-fH67*1rq+)DlI*nM?b?Nmpm7&`Yz!8v0U
zL)SdAthJr3wU`J~*AYhH$^MrWsm=VX?z0fR%|tKQ!UmcU0}eYAau~r7nvjMhBtXDL
zB|5_q=Fou}R3Q!lu)$d&nnDlCP*BK4c`QNr{q<n2xh@AMRgw3_hZ)ypK>AjD?-$V4
BHzoi8

diff --git a/egs/librispeech/ASR/conformer_ctc2/train.py b/egs/librispeech/ASR/conformer_ctc2/train.py
index aab2d2acb..146d10017 100755
--- a/egs/librispeech/ASR/conformer_ctc2/train.py
+++ b/egs/librispeech/ASR/conformer_ctc2/train.py
@@ -1066,9 +1066,9 @@ def run(rank, world_size, args):
             2**22
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)
-
+    
+    '''
     tedlium = TedLiumAsrDataModule(args)
-
     train_cuts = tedlium.train_cuts()
 
     if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
@@ -1084,6 +1084,54 @@ def run(rank, world_size, args):
 
     valid_cuts = tedlium.dev_cuts()
     valid_dl = tedlium.valid_dataloaders(valid_cuts)
+    '''
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    if params.full_libri:
+        train_cuts = librispeech.train_all_shuf_cuts()
+    else:
+        train_cuts = librispeech.train_clean_100_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 20 seconds
+        #    
+        # Caution: There is a reason to select 20.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #    
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 30.0 
+
+    def remove_invalid_utt_ctc(c: Cut):
+        # Caution: We assume the subsampling factor is 4!
+        # num_tokens = len(sp.encode(c.supervisions[0].text, out_type=int))
+        num_tokens = len(graph_compiler.texts_to_ids(c.supervisions[0].text))
+        min_output_input_ratio = 0.0005
+        max_output_input_ratio = 0.1
+        return (
+            min_output_input_ratio
+            < num_tokens / float(c.features.num_frames)
+            < max_output_input_ratio
+        )    
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+    train_cuts = train_cuts.filter(remove_invalid_utt_ctc)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None 
+
+    train_dl = librispeech.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )    
+
+    valid_cuts = librispeech.dev_clean_cuts()
+    valid_cuts += librispeech.dev_other_cuts()
+    valid_dl = librispeech.valid_dataloaders(valid_cuts)
 
     if (
         params.start_epoch <= 1