blob: 350c742329b82b496de277d5a5f2bbb711bc4125 (
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
|
This is a simple (and parallelizable) tuning method for cdec
which is able to train the weights of very many (sparse) features.
It was used here:
"Joint Feature Selection in Distributed Stochastic
Learning for Large-Scale Discriminative Training in
SMT"
(Simianer, Riezler, Dyer; ACL 2012)
Building
--------
Builds when building cdec, see ../BUILDING .
To build only parts needed for dtrain do
```
autoreconf -ifv
./configure [--disable-gtest]
cd dtrain/; make
```
Running
-------
To run this on a dev set locally:
```
#define DTRAIN_LOCAL
```
otherwise remove that line or undef, then recompile. You need a single
grammar file or input annotated with per-sentence grammars (psg) as you
would use with cdec. Additionally you need to give dtrain a file with
references (--refs) when running locally.
The input for use with hadoop streaming looks like this:
```
<sid>\t<source>\t<ref>\t<grammar rules separated by \t>
```
To convert a psg to this format you need to replace all "\n"
by "\t". Make sure there are no tabs in your data.
For an example of local usage (with the 'distributed' format)
the see test/example/ . This expects dtrain to be built without
DTRAIN_LOCAL.
Next
----
+ approx. Bleu?
+ turn off inclusion
+ (dtrain|decoder) more meta-parameters testing
+ feature selection directly in dtrain
+ feature template: target side rule ngrams
+ sa-extract -> leave-one-out for grammar of training set?
+ make svm doable; no subgradient?
+ reranking while sgd?
+ try PRO, mira emulations
+ sampling (MBR)
+ forest (on train)?
+ best BLEU transl (Sokolov)?
+ entire reg. path
+ resharding [nfold cross val.]
+ bigger LM, feats (target side Ng., word alignments etc.)
+ merge kbest lists
Legal
-----
Copyright (c) 2012 by Patrick Simianer <p@simianer.de>
See the file ../LICENSE.txt for the licensing terms that this software is
released under.
|
