blob: 8c328b7a9bd2940c0c83eccc6eef583f70b4dd66 (
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
|
cdec option "promise_power"
po::value<double>()->default_value(0)
Give more beam budget to more promising previous-pass nodes when pruning - but
allocate the same average beams. 0 means off, 1 means beam proportional to
inside*outside prob, n means nth power (affects just --cubepruning_pop_limit).
note: for the same poplimit, this gives more search error unless very close to 0
(recommend disabled; even 0.01 is slightly worse than 0) which is a bad sign and
suggests this isn't doing a good job; further it's slightly slower to LM cube
rescore with 0.01 compared to 0, as well as giving (very insignificantly) lower
BLEU. TODO: test under more conditions, or try idea with different formula, or
prob. cube beams.
with urdu mt09 2pass 1st pass tuned w/ unigram, prelm_beam_prune=1 scale_prune_srclen
promise_power: BLEU rescoring time
0: 20.21 1.8s
0.01: 20.16 1.9s
0.1: 20.11 2.0s
(insignificant differences)
10: 19.7 1.1s
(faster but much worse bleu)
|
