“2013-04-19”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
400 hour DNN training
 
(2位用户的27个中间修订版本未显示)
第1行: 第1行:
1. Data sharing
+
==Data sharing==
  
(1) AM/lexicon/LM are shared.
+
* AM/lexicon/LM are shared.
 +
* LM count files are still in transfering.  
  
(2) LM count files are still in transfering.  
+
==DNN progress==
 +
===400 hour DNN training===
 +
{| class="wikitable"
 +
!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
  
2. DNN progress
+
===Tencent test result===
  
  (1) 400 hour BN model.  
+
: 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
  
(2) Tencent test result: 70h training data(2 day, 15 machines, 10 threads),
 
88k LM, general test case:
 
  
gmmi-bmmi: 38.7%
+
===GPU & CPU merge===
 +
: Invesigate the possibility to merge GPU and CPU code. Try to find out an easier way. (1 week)
  
dnn-1: 28% 11 frame window,  phone-based tree
+
===L-1 sparse initial training===
 +
Start to investigating.
  
dnn-2: 34%  9  frame window,  state-based tree
+
==Kaldi/HTK merge==
 +
:* HTK2Kaldi: the tool with Kaldi does not work.
 +
:* Kaldi2HTK: done with implementation. Testing?
  
(3) GPU & CPU merge. Invesigate the possibility to merge GPU and
+
==Embedded progress==
CPU code. Try to find out an easier way. (1 week)
+
:* Some large performance (speed) degradation with the embedded platform(1/60).
 
+
:* Planning for sparse DNN.
(4) L-1 sparse initial training.
+
:* QA LM training, still failed. Mengyuan need more work on this.
 
+
3.Kaldi/HTK merge
+
 
+
(1) HTK2Kaldi: the tool with Kaldi does not work.
+
(2) Kaldi2HTK: done with implementation. Testing?
+
 
+
4. Embedded progress
+
 
+
(1). Some large performance (speed) degradation with the embedded platform(1/60).
+
 
+
(2). Planning for sparse DNN.
+
 
+
(3). 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.