summaryrefslogtreecommitdiff
path: root/dtrain/README.md
blob: 46f783b0cedfd459ea700c9a38e49aca695a5ad2 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
dtrain
======

Build & run
-----------
build ..
<pre>
git clone git://github.com/qlt/cdec-dtrain.git
cd cdec-dtrain
autoreconf -if[v]
./configure [--disable-gtest]
make
</pre>
and run:
<pre>
cd dtrain/hstreaming/
(edit ini files)
edit the vars in hadoop-streaming-job.sh ($ID, $IN and $OUT)
./hadoop-streaming-job.sh
</pre>

Ideas
-----
* *MULTIPARTITE* ranking (1 vs rest, cluster model/score)
* *REMEMBER* sampled translations (merge kbest lists)
* *SELECT* iteration with highest real BLEU on devtest?
* *GENERATED* data? (perfect translation always in kbest)
* *CACHE* ngrams for scoring
* hadoop *PIPES* imlementation
* *ITERATION* variants (shuffle resulting weights, re-iterate)
* *MORE THAN ONE* reference for BLEU?
* *RANDOM RESTARTS* or directions
* use separate *TEST SET* for each shard
* *REDUCE* training set (50k?)
* *SYNTAX* features (CD)
* distribute *DEV* set to all nodes, avg
* *PARAPHRASES* for better approx BLEU?


Uncertain, known bugs, problems
-------------------------------
* cdec kbest vs 1best (no -k param), rescoring (ref?)? => ok(?)
* no sparse vector in decoder => ok/fixed
* PhraseModel features, mapping?
* flex scanner jams on bad input, we could skip that
* input/grammar caching (strings -> WordIDs)
* look at forest sampling...
* devtest loo or not? why loo grammars larger? (sort psgs | uniq -> grammar)
* lower beam size to be faster?
* why is <unk> -100 in lm so good?
* noise helps for discriminative training?
* what does srilm do with -unk but nothing mapped to unk (<unk> unigram)?
  => this: http://www-speech.sri.com/pipermail/srilm-user/2007q4/000543.html
* mira translation sampling? => done
* does AER correlate with BLEU?

random notes
------------
* learning rate tuned with perceptron
* aer correlation with bleu?
* dtrain (perc) used for some tests because no optimizer instability
* http://www.ark.cs.cmu.edu/cdyer/dtrain/
* repeat as often as max needed by any learner!
* don't compare lms with diff vocab (stupid backoff paper)
* what does mira/pro optimize?
* early stopping
* 10-20k rules per sent normal
* shard size 500 -> 2k
* giza vs. berkeleyaligner: giza less noise?
* compound splitting -> more rules?
* loo => ref can't be reached? (jackknifing)
* prune singletons -> less noise? (do I do this?)
* random sample: take 100 at random

features
--------
* baseline features (take whatever cdec implements for VEST)
* rule identifiers (feature name = rule as string)
* rule discounts (taken from frequency i or frequency interval [i,j] of rule in extraction from parallel training data) bins
* target ngrams (from nonterminals in rule rhs), with gaps?
* source-target unigrams (from word alignments used in rule extraction, if they are?)
* lhs, rhs, rule length features
* all other features depend on syntax annotation. 
* word alignment

FIXME, todo
-----------
* merge dtrain part-X files, for better blocks (how to do this with 4.5tb ep)
* mapred count shard sents
* mapred stats for learning curve (output weights per iter for eval on devtest)
* 250 forest sampling is real bad, bug?
* metric reporter of bleu for each shard (reporters, status?)
  to draw learning curves for all shards in 1 plot
* kenlm not portable (i7-2620M vs Intel(R) Xeon(R) CPU E5620 @ 2.40GHz)
* mapred chaining? hamake?
* make our sigtest work with cdec
* l1l2 red (tsuroke)?
* epsilon stopping criterion
* normalize weight vector to get proper model scores for forest sampling
* 108010 with gap(s), and/or fix (same score in diff groups)
* 108010: combine model score + bleu
* visualize weight vector
* *100 runs stats
* correlation of *_bleu to ibm_bleu
* ep: open lm, cutoff @1
* tune regs
* 3x3 4x4 5x5 .. 10x10 until standard dev ok, moving avg
* avg weight vector for dtrain? (mira non-avg)
* repeat lm choose with mira/pro
* shuffle training data
* learning rate dynamic (Duh? Tsuroka?)
* divide updates by ?
* mira: 5/10/15, pro: (5)/10/20/30 (on devtest!)
* sample pairs like in pro
* mira forest sampling


Data
----
<pre>
nc-v6.de-en             apegd
nc-v6.de-en.loo         apegd
nc-v6.de-en.giza        apegd
nc-v6.de-en.giza.loo    apegd
nc-v6.de-en.cs.giza     apegd
nc-v6.de-en.cs.giza.loo apegd
nv-v6.de-en.cs          apegd
nc-v6.de-en.cs.loo      apegd
--
ep-v6.de-en.cs          apegd
ep-v6.de-en.cs.loo      apegd

a: alignment:, p: prep, e: extract,
g: grammar, d: dtrain
</pre>

Experiments
-----------
[grammar stats
  oov on dev/devtest/test
  size
  #rules (uniq)
  time for building
   ep: 1.5 days on 278 slots (30 nodes)
   nc: ~2 hours ^^^

 lm stats
  oov on dev/devtest/test 
  perplex on train/dev/devtest/test?]

[0]
which word alignment?
 berkeleyaligner
 giza++ as of Sep 24 2011, mgizapp 0.6.3
 --symgiza as of Oct 1 2011--
 ---
 NON LOO
 (symgiza unreliable)
 randomly sample 100 from train with loo
 run dtrain for 100 iterations
 w/o all other feats (lm, wp, ...) +Glue
 measure ibm bleu on exact same sents
 ep -> berkeleyaligner ??? (mb per sent, rules per sent)

*100 -> triples, quadruples

[1]
lm?
 3-4-5
 open
 unk
 nounk (-100 for unk)
 --
 lm oov weight pos? -100
 no tuning, -100 prob for unk EXPECT: nounk
 tuning with dtrain EXPECT: open
 =>
  lmtest on cs.giza.loo???

[2]
cs?
 'default' weights

[3]
loo vs non-loo
 'jackknifing'
 generalization (determ.!) on dev, test on devtest

[4]
stability
 all with default params
 mira: 100
 pro: 100
 vest: 100
 dtrain: 100

[undecided]
do we even need loo for ep?
pro metaparam
 (max) iter
 regularization
 ???
 
mira metaparam
 (max) iter: 10 (nc???) vs 15 (ep???)

features to try
 NgramFeatures -> target side ngrams
 RuleIdentityFeatures
 RuleNgramFeatures -> source side ngrams from rule
 RuleShape -> relative orientation of X's and terminals
 SpanFeatures -> http://www.cs.cmu.edu/~cdyer/wmt11-sysdesc.pdf
 ArityPenalty -> Arity=0 Arity=1 and Arity=2

---
shard size: 500-2k
iterations, re-iterate (shuffle w): 10
gamma, eta
SVM, perceptron
reducer: avg (feats/shard), l1l2, active on all shards
sentence sampling: forest
pair sampling: all, rand, 108010 (sort), PRO
out of domain test?

---
variables to control

[alignment]

[lm]

[vest]

[mira]

[dtrain]

[pro]


--------
In PRO, a continually growing list of candidates is maintained for
each sentence by concatenating k-best lists from each decoding run,
and the training pairs are sampled from them. This is done to ensure
that the optimizer doesn't forget about bad places in the parameter
space that it visited previously (since some training samples will be
selected from that space). Something like your approach should work
well though, provided you don't overfit to the sentence pair you're
looking at in each iteration. So I guess the question is: what are you
doing in step 2 exactly? A complete optimization? Taking one step? The
other thing is, do you maintain n-best hypotheses from previous
iterations?

--------
good grammar? => ability to overfit
 berkeley vs giza
 not LOO
 NO optimizer instability
 20+ iterations
 approx_bleu-4
 train on dev => test on dev
 train on devtest => test on devtest
 dev on dev better?
 devtest on devtest better?
 (train/test on loo? => lower!)
 (test on others => real bad)


loo vs non-loo? => generalization
 (cs vs non-cs?)
 giza||berkeley
 LOO + non LOO
 2 fold cross validation 
 train on dev, test on devtest
 train on devtest, test on dev
 as above ^^^