Hi,

we're experiencing random segfaults when we use two phrase tables in 
alternative decoding paths. The exact commit of moses we use is 
6a06e7776a58b09e4ed5b1cf11eb64fbdd6b02a2, from April 1.

We do have test runs on the exact same 200 input sentences, exact same 
moses.ini, on the very same machine, where one of the runs succeeds and the 
other dies after 45 sentences.


Would anyone have any idea what should we be chasing?

- it doesn't seem to be thread-related (segfault experienced with -threads 1 as 
well as -threads 8)
- not related to nbest-list construction (we first had this problem in mert 
tuning so we isolated this)
- not related to more LMs (we first had several LMs in the setup, we get the 
crash with just one as well)
- not related to -search, the bug is there with -search set to 0, 1 or 4
- seems related to data or data size: when we trained the first ttable on just 
a very small corpus, we did not get the segfault (yet)
- not related to translation options caching, the bug is there even with 
-no-cache
- not related to the specification of output-factors; left unspecified or set 
to 0<CR>1, the bug is there


Here is the moses.ini:

[input-factors]
0

[mapping]
0 T 0
1 T 1

[distortion-limit]
6

[feature]
Distortion
KENLM lazyken=0 name=LM0 factor=0 path=lm.1.trie.lm order=4
PhraseDictionaryCompact name=TranslationModel0 num-features=4 
path=phrase-table.0-0,1.1.1 input-factor=0 output-factor=0,1 table-limit=100
PhraseDictionaryCompact name=TranslationModel1 num-features=4 
path=phrase-table.0-0,1.2.1 input-factor=0 output-factor=0,1 table-limit=100
PhrasePenalty
UnknownWordPenalty
WordPenalty

[weight]
Distortion0= 0.3
LM0= 0.5
PhrasePenalty0= 0.2
TranslationModel0= 0.2 0.2 0.2 0.2
TranslationModel1= 0.2 0.2 0.2 0.2
UnknownWordPenalty0= 1
WordPenalty0= -1


The large setup that shows these crashes uses this big files:

-rw-r--r-- 1 bojar ufal 584M Apr 13 09:19 lm.1.trie.lm
-rw-r--r-- 1 bojar ufal 1.1G Apr 13 09:24 phrase-table.0-0,1.1.1.minphr
-rw-r--r-- 1 bojar ufal 5.7M Apr 13 09:24 phrase-table.0-0,1.2.1.minphr


Thanks,
  Ondrej.


-- 
Ondrej Bojar (mailto:[email protected] / [email protected])
http://www.cuni.cz/~obo
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to