“2013-04-19”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
400 hour DNN training
 
(1位用户的12个中间修订版本未显示)
第5行: 第5行:
  
 
==DNN progress==
 
==DNN progress==
===400 hour BN model ===
+
===400 hour DNN training===
# Tencent baseline:
+
{| class="wikitable"
<pre>
+
!Test Set!! Tencent Baseline!! bMMI!! fMMI !! BN(with fMMI) !! Hybrid
700 hour online data + 700 863 data , HLDA+MPE; 88k lexicon:
+
|-
 +
|1900||8.4  || 7.65 || 7.35||6.57 || 7.27
 +
|-
 +
|2044|| 22.4 ||24.44|| 24.03||21.77 || 20.24
 +
|-
 +
|online1||35.6 ||34.66||34.33||31.44 || 30.53
 +
|-
 +
|online2||29.6 ||27.23||26.80||24.10 || 23.89
 +
|-
 +
|map||24.5|| 27.54||27.69||23.79 || 22.46
 +
|-
 +
|notepad||16|| 19.81||21.75||15.81 || 12.74
 +
|-
 +
|general||36|| 38.52||38.90||33.61 || 31.55
 +
|-
 +
|speedup||26.8||27.88||26.81||22.82 || 22.00
 +
|-
 +
|}
 +
*Tencent baseline is with 700h online data+ 700h 863 data, HLDA+MPE, 88k lexicon
 +
*Our results are with 400 hour AM, 88k LM. ML+bMMI
  
record1900: 8.4
 
2044:      22.4
 
online 1:  35.6
 
online 2:  29.6
 
map:        24.5
 
notepad:    16
 
general:    36
 
speedup:    26.8
 
</pre>
 
# bMMI
 
<pre>
 
exp/tri4b_mmi_b0.1/decode_tlm_biglm:
 
map: %WER 27.54 [ 4029 / 14628, 63 ins, 533 del, 3433 sub ]
 
2044: %WER 24.44 [ 5681 / 23241, 313 ins, 844 del, 4524 sub ]
 
notetp3: %WER 19.81 [ 367 / 1853, 8 ins, 48 del, 311 sub ]
 
record1900: %WER 7.65 [ 909 / 11888, 17 ins, 377 del, 515 sub ]
 
general: %WER 38.52 [ 14490 / 37619, 182 ins, 1314 del, 12994 sub ]
 
online1: %WER 34.66 [ 9855 / 28433, 398 ins, 1895 del, 7562 sub ]
 
online2: %WER 27.23 [ 16092 / 59101, 623 ins, 2954 del, 12515 sub ]
 
speedup: %WER 27.88 [ 1465 / 5255, 32 ins, 332 del, 1101 sub ]
 
</pre>
 
#fMMI
 
<pre>
 
exp/tri4b_fmmi_indirect/decode_tlm_it7_biglm:
 
map: %WER 27.69 [ 4050 / 14628, 61 ins, 538 del, 3451 sub ]
 
2044: %WER 24.03 [ 5584 / 23241, 316 ins, 817 del, 4451 sub ]
 
notetp3: %WER 21.75 [ 403 / 1853, 7 ins, 53 del, 343 sub ]
 
record1900: %WER 7.35 [ 874 / 11888, 31 ins, 347 del, 496 sub ]
 
general: %WER 38.90 [ 14635 / 37619, 206 ins, 1331 del, 13098 sub ]
 
online1: %WER 34.33 [ 9762 / 28433, 424 ins, 1888 del, 7450 sub ]
 
online2: %WER 26.80 [ 15837 / 59101, 648 ins, 2902 del, 12287 sub ]
 
speedup: %WER 26.81 [ 1409 / 5255, 35 ins, 284 del, 1090 sub ]
 
</pre>
 
#DNN-bn
 
<pre>
 
exp/tri4d_fmmi_indirect/decode_tlm_it4_biglm:
 
map: %WER 23.79 [ 3480 / 14628, 58 ins, 465 del, 2957 sub ]
 
2044: %WER 21.77 [ 5060 / 23241, 297 ins, 711 del, 4052 sub ]
 
notetp3: %WER 15.81 [ 293 / 1853, 8 ins, 35 del, 250 sub ]
 
record1900: %WER 6.57 [ 781 / 11888, 18 ins, 325 del, 438 sub ]
 
general: %WER 33.61 [ 12645 / 37619, 191 ins, 968 del, 11486 sub ]
 
online1: %WER 31.44 [ 8940 / 28433, 311 ins, 1619 del, 7010 sub ]
 
online2: %WER 24.10 [ 14245 / 59101, 523 ins, 2417 del, 11305 sub ]
 
speedup: %WER 22.82 [ 1199 / 5255, 39 ins, 241 del, 919 sub ]
 
</pre>
 
 
===Tencent test result===
 
===Tencent test result===
  
第72行: 第46行:
  
 
==Kaldi/HTK merge==
 
==Kaldi/HTK merge==
#: HTK2Kaldi: the tool with Kaldi does not work.
+
:* HTK2Kaldi: the tool with Kaldi does not work.
#: Kaldi2HTK: done with implementation. Testing?
+
:* Kaldi2HTK: done with implementation. Testing?
  
 
==Embedded progress==
 
==Embedded progress==
#: Some large performance (speed) degradation with the embedded platform(1/60).
+
:* Some large performance (speed) degradation with the embedded platform(1/60).
#: Planning for sparse DNN.
+
:* Planning for sparse DNN.
#: QA LM training, still failed. Mengyuan need more work on this.
+
:* QA LM training, still failed. Mengyuan need more work on this.

2013年4月26日 (五) 05:31的最后版本

Data sharing

  • AM/lexicon/LM are shared.
  • LM count files are still in transfering.

DNN progress

400 hour DNN training

Test Set Tencent Baseline bMMI fMMI BN(with fMMI) Hybrid
1900 8.4 7.65 7.35 6.57 7.27
2044 22.4 24.44 24.03 21.77 20.24
online1 35.6 34.66 34.33 31.44 30.53
online2 29.6 27.23 26.80 24.10 23.89
map 24.5 27.54 27.69 23.79 22.46
notepad 16 19.81 21.75 15.81 12.74
general 36 38.52 38.90 33.61 31.55
speedup 26.8 27.88 26.81 22.82 22.00
  • Tencent baseline is with 700h online data+ 700h 863 data, HLDA+MPE, 88k lexicon
  • Our results are with 400 hour AM, 88k LM. ML+bMMI

Tencent test result

AM: 70h training data(2 day, 15 machines, 10 threads)
LM: 88k LM
Test case: general
gmmi-bmmi: 38.7%
dnn-1: 28% 11 frame window, phone-based tree
dnn-2: 34% 9 frame window, state-based tree


GPU & CPU merge

Invesigate the possibility to merge GPU and CPU code. Try to find out an easier way. (1 week)

L-1 sparse initial training

Start to investigating.

Kaldi/HTK merge

  • HTK2Kaldi: the tool with Kaldi does not work.
  • Kaldi2HTK: done with implementation. Testing?

Embedded progress

  • Some large performance (speed) degradation with the embedded platform(1/60).
  • Planning for sparse DNN.
  • QA LM training, still failed. Mengyuan need more work on this.