“Asr-language-processing-research-rnng-mn”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
（→‎Time Table）
（→‎Time Table）
 
（相同用户的12个中间修订版本未显示）
第8行： 第8行：
 
!Date !! Work Plan !! Work Done !! Completion Rate
 
!Date !! Work Plan !! Work Done !! Completion Rate
 
|-
 
|-
−
| rowspan="6"|2016/10/31
+
| 2016/10/31
−
| ||
+
|  
 +
* implement rnng+static memory discriminative model
 +
# fix the unexpected action
 +
#  rerun the original discriminative model
 +
# rerun the centred memory rnng model
 +
#  get wrong instances of original trained model, and get statistics
 +
# run the wrong memory rnng model
 +
# run the sampled memory rnng model
 +
#  update experiment report
 
||
 
||
−
|-
+
* implementation is done, but result is not satisfied.
−
| ||
+
#  fixed the unexpected action
−
 
+
# reran the original discriminative model
 +
#  reran the centred memory rnng model
 +
#  got wrong instances of original trained model, and get statistics
 +
# ran the sampled memory rnng model
 +
#  ran the wrong memory rnng model
 +
# updated experiment report [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/2/2f/RNNG%2Bmm%E5%AE%9E%E9%AA%8C%E6%8A%A5%E5%91%8A.pdf]
 
||
 
||
−
 
+
* 100%
 
|-
 
|-
−
| ||  
+
| 2016/11/7
−
 
+
|  
 +
* modify discriminative model,  try to prove the positive function of static memory
 +
# change train set
 +
# add parameter before cos
 +
# modify model structure
 +
* read the code of Teacher Feng
 +
# understand and run
 +
* how to use GPU in rnng
 +
# learn to use it
 +
|| 
 +
* modify discriminative model,  try to prove the positive function of static memory
 +
# changed train set
 +
# added parameter before cos
 +
# modified model structure
 +
* read the code of Teacher Feng
 +
# understood but did't run
 +
* how to use GPU in rnng
 +
# the result is rnng cannot run fast on GPU
 +
* update the report [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/2/2f/RNNG%2Bmm%E5%AE%9E%E9%AA%8C%E6%8A%A5%E5%91%8A.pdf report]
 
||
 
||
−
 
+
* 90%
 
|-
 
|-
−
| ||
+
| 2016/11/14
−
 
+
|
−
||
+
* try to run rnng on multi cpu cores
−
 
+
* refine the models tried before and give the result report
 +
* finish the code of dynamic memory model
 +
||
 +
* implemented on MKL successfully
 +
* reran original model
 +
* reran the models  tried before, but still running
 +
* finished the code of dynamic memory model, ran and got a result
 +
* tried another structure of memory
 +
||
 +
* 90%
 
|-
 
|-
−
| ||
+
| 2016/11/21
−
 
+
|  
−
||
+
* try more different models
−
 
+
# rerun original model, if the same result
 +
# run more dynamic memory models, hope to see better results
 +
# run models with another structure of memory, if it is better than previous structure
 +
* summary experiments' results and give report
 +
* publish the TRPs of RNNG
 +
|| ||
 
|-
 
|-
−
| ||  
+
| 2016/11/28
−
 
+
|
−
||
+
* modify model
−
 
+
* try to prove the positive function of dynamic memory
 +
||  ||
 +
|-
 +
| 2016/12/5
 +
|  
 +
* get the first final rnng+mm discriminative model
 +
||  ||
 +
|-
 +
| 2016/12/12
 +
|
 +
* give a plan to transfer to generative model
 +
||  ||
 +
|-
 +
| 2016/12/19
 +
|
 +
* implement  rnng+mm generative model
 +
||  ||
 +
|-
 +
| 2016/12/26
 +
|
 +
* modify model
 +
* try to prove the positive function of memory
 +
||  ||
 
|-
 
|-
 
|}
 
|}
  
 
==Progress==
 
==Progress==

2016年11月21日 (一) 06:55的最后版本

Main Idea

People

Yang Feng, Shiyue Zhang, Andi Zhang

Time Table

Date Work Plan Work Done Completion Rate
2016/10/31
  • implement rnng+static memory discriminative model
  1. fix the unexpected action
  2. rerun the original discriminative model
  3. rerun the centred memory rnng model
  4. get wrong instances of original trained model, and get statistics
  5. run the wrong memory rnng model
  6. run the sampled memory rnng model
  7. update experiment report
  • implementation is done, but result is not satisfied.
  1. fixed the unexpected action
  2. reran the original discriminative model
  3. reran the centred memory rnng model
  4. got wrong instances of original trained model, and get statistics
  5. ran the sampled memory rnng model
  6. ran the wrong memory rnng model
  7. updated experiment report [1]
  • 100%
2016/11/7
  • modify discriminative model, try to prove the positive function of static memory
  1. change train set
  2. add parameter before cos
  3. modify model structure
  • read the code of Teacher Feng
  1. understand and run
  • how to use GPU in rnng
  1. learn to use it
  • modify discriminative model, try to prove the positive function of static memory
  1. changed train set
  2. added parameter before cos
  3. modified model structure
  • read the code of Teacher Feng
  1. understood but did't run
  • how to use GPU in rnng
  1. the result is rnng cannot run fast on GPU
  • 90%
2016/11/14
  • try to run rnng on multi cpu cores
  • refine the models tried before and give the result report
  • finish the code of dynamic memory model
  • implemented on MKL successfully
  • reran original model
  • reran the models tried before, but still running
  • finished the code of dynamic memory model, ran and got a result
  • tried another structure of memory
  • 90%
2016/11/21
  • try more different models
  1. rerun original model, if the same result
  2. run more dynamic memory models, hope to see better results
  3. run models with another structure of memory, if it is better than previous structure
  • summary experiments' results and give report
  • publish the TRPs of RNNG
2016/11/28
  • modify model
  • try to prove the positive function of dynamic memory
2016/12/5
  • get the first final rnng+mm discriminative model
2016/12/12
  • give a plan to transfer to generative model
2016/12/19
  • implement rnng+mm generative model
2016/12/26
  • modify model
  • try to prove the positive function of memory

Progress

取自“http://index.cslt.org/mediawiki/index.php?title=Asr-language-processing-research-rnng-mn&oldid=23973”