“2013-06-07”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
（以内容“== Data sharing == * LM count files still undelivered! == DNN progress == === Experiments === * sparse DNN: sticky training (retrain the nnet while keeping the spars...”创建新页面）
 
 
（相同用户的4个中间修订版本未显示）
第21行： 第21行：
 
|}
 
|}
  
−
The conclusion is that with the L2 retrain, the DNN performance is largely called back. The extremely sparse case (th0.3) with sticky training seems quite amazing. This means the network could be sparse. However this is just for the 1900 test. Need test on other sets.
+
Conclusion: The extremely sparse network can largely pertain the performance of DNN. The structure seems more important than the parameter tuning based on the structure.  
  
 
* fixed-point DNN forwarding
 
* fixed-point DNN forwarding
  
−
According to the fixed-point FST and NN, and the results of the sparse NN, we are working on fast NN decoder which is suitable for embedded device. The work is just started.
+
#working on migrating the Atlas lib to ARM.
 +
#working on atlas/mkl independent implementation.
  
 
=== Tencent exps ===
 
=== Tencent exps ===
  
−
本周1000小时实验已结束，实验性能如下：
+
*6000h model training, could be finished on 25th approximately.
 +
*working on sequential DNN DT: refer to "Error Back Propagation For Sequence Training of Context-Dependent Deep Networks For Conversation Speech Transcription"
  
−
{| class="wikitable"
 
−
!!!  old baseline !!  new baseline !!  DNN
 
−
|-
 
−
|1900  ||  8.4 || 6.8 || 4.3
 
−
|-
 
−
|2044  ||  22.4  || 15.7  || 12.7
 
−
|-
 
−
|online1 || 35.6 ||32.7 || 25.8
 
−
|-
 
−
|online2 || 29.6 ||27.3 || 22.1
 
−
|-
 
−
|map || 24.5 || 15.8 || 13.4
 
−
|-
 
−
|notepad || 16 || 8.1 || 5.6
 
−
|-
 
−
|general || 36 || 25.1 || 19.3
 
−
|-
 
−
|speedup || 26.8 || 14 || -
 
−
|-
 
−
|}
 
−
 
−
接下来计划：
 
−
*6000小时模型训练，dnn模型相关其他技术（序列化dt，alignment，pretrain）
 
  
 
=== GPU & CPU merge ===
 
=== GPU & CPU merge ===
第59行： 第38行：
 
#on progress.
 
#on progress.
  
−
 
+
== RNN LM progress ==
−
== Kaldi/HTK merge ==
+
*Initial work started. 100M data with a 10k vocabulary obtained a perplexity 180.
−
 
+
*More exploration continuous.  
−
* HTK2Kaldi: hold.
+
−
* Kaldi2HTK: hold and second priority
+
−
 
+
−
The above work is probably not very necessary since Tencent will fully migrate to the hybrid DNN approach, and therefore HTK will be never used.
+
  
 
== Embedded progress ==
 
== Embedded progress ==
第71行： 第46行：
 
*Status:
 
*Status:
 
: check the reference, and change the compiling options
 
: check the reference, and change the compiling options
−
: the large-scale AM training based on the Tencent 400h data is done.
+
: the large-scale AM training based on the Tencent 400h data is done, continuous HMM.
−
: the random output problem is fixed.
+
 
  
 
{| class="wikitable"
 
{| class="wikitable"
−
! Test Set !! #utt !! PS default !! Tencent
+
! Sys !! WER !! RT
 
|-
 
|-
−
| cw  || 993 || 8.01(RT: 0.07) || 7.61(RT: 0.40)
+
| SP model      || 8.81  || 0.07
 
|-
 
|-
−
| hfc || 986 || 6.69(RT: 0.07) || 5.48(RT: 0.40)
+
| Tencent tone  || 6.33  || 0.40
−
|-  
+
|-
−
| zz  || 984 || 12.73(RT: 0.07) || 5.91(RT: 0.40)
+
| Tencent notone || 5.04  || 0.31
 
|-
 
|-
 
|}
 
|}
−
 
  
 
*To be done
 
*To be done
 
:# large scale parallel training.
 
:# large scale parallel training.
 
:# NN based engine(dynamic and static).
 
:# NN based engine(dynamic and static).
 +
:# Semi-continuous model with the Tencent data
 +
:# Debug on external an ARM board.

2013年6月7日 (五) 08:40的最后版本

Data sharing

  • LM count files still undelivered!

DNN progress

Experiments

  • sparse DNN: sticky training (retrain the nnet while keeping the sparsness)

zero small values(test set: 1900), with extremely sparseness:

threshold 0 0.2 0.3 0.4 0.5
shrinkage% 0.0 66.4 81.6 0.90 0.94
without sticky: WER 7.55 9.46 53.23 98.99 -
with sticky: WER 7.55 7.56 7.87 8.81 9.87

Conclusion: The extremely sparse network can largely pertain the performance of DNN. The structure seems more important than the parameter tuning based on the structure.

  • fixed-point DNN forwarding
  1. working on migrating the Atlas lib to ARM.
  2. working on atlas/mkl independent implementation.

Tencent exps

  • 6000h model training, could be finished on 25th approximately.
  • working on sequential DNN DT: refer to "Error Back Propagation For Sequence Training of Context-Dependent Deep Networks For Conversation Speech Transcription"


GPU & CPU merge

  1. on progress.

RNN LM progress

  • Initial work started. 100M data with a 10k vocabulary obtained a perplexity 180.
  • More exploration continuous.

Embedded progress

  • Status:
check the reference, and change the compiling options
the large-scale AM training based on the Tencent 400h data is done, continuous HMM.


Sys WER RT
SP model 8.81 0.07
Tencent tone 6.33 0.40
Tencent notone 5.04 0.31
  • To be done
  1. large scale parallel training.
  2. NN based engine(dynamic and static).
  3. Semi-continuous model with the Tencent data
  4. Debug on external an ARM board.
取自“http://cslt.org/mediawiki/index.php?title=2013-06-07&oldid=7842”