Difference between revisions of "Schedule"

From cslt Wiki
Jump to: navigation, search
(Daily Report)
(Daily Report)
Line 330: Line 330:
 
   force teaching mechanism (training process) and beam search mechanism (decoding process)
 
   force teaching mechanism (training process) and beam search mechanism (decoding process)
 
   propagates and expands the error to the output end, which destroys the model when decoding.
 
   propagates and expands the error to the output end, which destroys the model when decoding.
 +
*next:
 +
  Try to train double-decoder model without joint loss but with beam search on 1st decoder.
 
|-
 
|-
 
|}
 
|}

Revision as of 09:12, 23 May 2017

NLP Schedule

Members

Current Members

  • Yang Feng (冯洋)
  • Jiyuan Zhang (张记袁)
  • Aodong Li (李傲冬)
  • Andi Zhang (张安迪)
  • Shiyue Zhang (张诗悦)
  • Li Gu (古丽)
  • Peilun Xiao (肖培伦)
  • Shipan Ren (任师攀)

Former Members

  • Chao Xing (邢超)  : FreeNeb
  • Rong Liu (刘荣)  : 优酷
  • Xiaoxi Wang (王晓曦) : 图灵机器人
  • Xi Ma (马习)  : 清华大学研究生
  • Tianyi Luo (骆天一) : phd candidate in University of California Santa Cruz
  • Qixin Wang (王琪鑫)  : MA candidate in University of California
  • DongXu Zhang (张东旭): --
  • Yiqiao Pan (潘一桥) : MA candidate in University of Sydney
  • Shiyao Li (李诗瑶) : BUPT
  • Aiting Liu (刘艾婷)  : BUPT

Work Progress

Daily Report

Date Person start leave hours status
2017/04/02 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/03 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/04 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/05 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/06 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/07 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/08 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/09 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/10 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/11 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/12 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/13 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/14 Andy Zhang 9:30 18:30 8
  • preparing EMNLP
Peilun Xiao
2017/04/15 Andy Zhang 9:00 15:00 6
  • preparing EMNLP
Peilun Xiao
2017/04/18 Aodong Li 11:00 20:00 8
  • Pick up new task in news generation and do literature review
2017/04/19 Aodong Li 11:00 20:00 8
  • Literature review
2017/04/20 Aodong Li 12:00 20:00 8
  • Literature review
2017/04/21 Aodong Li 12:00 20:00 8
  • Literature review
2017/04/24 Aodong Li 11:00 20:00 8
  • Adjust literature review focus
2017/04/25 Aodong Li 11:00 20:00 8
  • Literature review
2017/04/26 Aodong Li 11:00 20:00 8
  • Literature review
2017/04/27 Aodong Li 11:00 20:00 8
  • Try to reproduce sc-lstm work
2017/04/28 Aodong Li 11:00 20:00 8
  • Transfer to new task in machine translation and do literature review
2017/04/30 Aodong Li 11:00 20:00 8
  • Literature review
2017/05/01 Aodong Li 11:00 20:00 8
  • Literature review
2017/05/02 Aodong Li 11:00 20:00 8
  • Literature review and code review
2017/05/06 Aodong Li 14:20 17:20 3
  • Code review
2017/05/07 Aodong Li 13:30 22:00 8
  • Code review and experiment started, but version discrepancy encountered
2017/05/08 Aodong Li 11:30 21:00 8
  • Code review and version discrepancy solved
2017/05/09 Aodong Li 13:00 22:00 9
  • Code review and experiment
  • details about experiment:
 small data, 
 1st and 2nd translator uses the same training data, 
 2nd translator uses random initialized embedding
  • results (BLEU):
 BASELINE: 43.87
 best result of our model: 42.56
2017/05/10 Shipan Ren 9:00 20:00 11
  • Entry procedures
  • Machine Translation paper reading
2017/05/10 Aodong Li 13:30 22:00 8
  • experiment setting:
 small data, 
 1st and 2nd translator uses the different training data, counting 22000 and 22017 seperately
 2nd translator uses random initialized embedding
  • results (BLEU):
 BASELINE: 36.67 (36.67 is the model at 4750 updates, but we use model at 3000 updates to
                    prevent the case of overfitting, to generate the 2nd translator's training data, for 
                    which the BLEU is 34.96)
 best result of our model: 29.81
 This may suggest that that using either the same training data with 1st translator or different
                   one won't influence 2nd translator's performance, instead, using the same one may
                    be better, at least from results. But I have to give a consideration of a smaller size 
                    of training data compared to yesterday's model.
  • code 2nd translator with constant embedding
2017/05/11 Shipan Ren 10:00 19:30 9.5
  • Configure environment
  • Run tf_translate code
  • Read Machine Translation paper
2017/05/11 Aodong Li 13:00 21:00 8
  • experiment setting:
 small data, 
 1st and 2nd translator uses the same training data, 
 2nd translator uses constant untrainable embedding imported from 1st translator's decoder
  • results (BLEU):
 BASELINE: 43.87
 best result of our model: 43.48
 Experiments show that this kind of series or cascade model will definitely impair the final perfor-
                     mance due to information loss as the information flows through the network from 
                     end to end. Decoder's smaller vocabulary size compared to encoder's demonstrate
                     this (9000+ -> 6000+).
 The intention of this experiment is looking for a map to solve meaning shift using 2nd translator,
                     but result of whether the map is learned or not is obscured by the smaller vocab size 
                     phenomenon.
  • literature review on hierarchical machine translation
2017/05/12 Aodong Li 13:00 21:00 8
  • Code double decoding model and read multilingual MT paper
2017/05/13 Shipan Ren 10:00 19:00 9
  • read machine translation paper
  • learne lstm model and seq2seq model
2017/05/14 Aodong Li 10:00 20:00 9
  • Code double decoding model and experiment
  • details about experiment:
 small data, 
 2nd translator uses as training data the concat(Chinese, machine translated English), 
 2nd translator uses random initialized embedding
  • results (BLEU):
 BASELINE: 43.87
 best result of our model: 43.53
  • NEXT: 2nd translator uses trained constant embedding
2017/05/15 Shipan Ren 9:30 19:00 9.5
  • understand the difference between lstm model and gru model
  • read the implement code of seq2seq model
2017/05/17 Shipan Ren 9:30 19:30 10
  • read neural machine translation paper
  • read tf_translate code
Aodong Li 13:30 24:00 9
  • code and debug double-decoder model
  • alter 2017/05/14 model's size and will try after nips
2017/05/18 Shipan Ren 10:00 19:00 9
  • read neural machine translation paper
  • read tf_translate code
Aodong Li 12:30 21:00 8
  • train double-decoder model on small data set but encounter decode bugs
2017/05/19 Aodong Li 12:30 20:30 8
  • debug double-decoder model
  • the model performs well on develop set, but performs badly on test data. I want to figure out the reason.
2017/05/21 Aodong Li 10:30 18:30 8
  • details about experiment:
 hidden_size = 700 (500 in prior)
 emb_size = 510 (310 in prior)
 small data, 
 2nd translator uses as training data the concat(Chinese, machine translated English), 
 2nd translator uses random initialized embedding
  • results (BLEU):
 BASELINE: 43.87
 best result of our model: 45.21
 But only one checkpoint outperforms the baseline, the other results are commonly under 43.1
  • debug double-decoder model
2017/05/22 Aodong Li 14:00 22:00 8
  • double-decoder without joint loss generalizes very bad
  • i'm trying double-decoder model with joint loss
2017/05/23 Aodong Li 13:00 22:00 8
  • details about experiment 1:
 hidden_size = 700
 emb_size = 510
 learning_rate = 0.0005 (0.001 in prior)
 small data, 
 2nd translator uses as training data the concat(Chinese, machine translated English), 
 2nd translator uses random initialized embedding
  • results (BLEU):
 BASELINE: 43.87
 best result of our model: 42.19
 Overfitting? In overall, the 2nd translator performs worse than baseline
  • details about experiment 2:
 hidden_size = 500
 emb_size = 310
 learning_rate = 0.001
 small data, 
 double-decoder model with joint loss which means the final loss  = 1st decoder's loss + 2nd 
 decoder's loss
  • results (BLEU):
 BASELINE: 43.87
 best result of our model: 39.04
 The 1st decoder's output is generally better than 2nd decoder's output. The reason may be that 
 the second decoder only learns from the first decoder's hidden states because their states are 
 almost the same.
  • DISCOVERY:
 The reason why double-decoder without joint loss generalizes very bad is that the gap between
 force teaching mechanism (training process) and beam search mechanism (decoding process)
 propagates and expands the error to the output end, which destroys the model when decoding.
  • next:
 Try to train double-decoder model without joint loss but with beam search on 1st decoder.

Time Off Table

Date Yang Feng Jiyuan Zhang

Past progress

nlp-progress 2017/03

nlp-progress 2017/02

nlp-progress 2017/01

nlp-progress 2016/12

nlp-progress 2016/11

nlp-progress 2016/10

nlp-progress 2016/09

nlp-progress 2016/08

nlp-progress 2016/05-07

nlp-progress 2016/04