Hi list, 

I had Moses segfault on me and traced the error back to "nan" being
returned by KenLM instead of a number, which then led to a crash during
hypothesis pruning. This happened with a binarized language model (trie)
that was trained on SRILM, and using a recent git version of Moses both
for LM binarization and decoding (0c88a1ec2e8c4). The non-binarized
version worked fine.

I have a reproducible test case, albeit with a 6 GB language model:

$ echo ", und den haben ." | ~/moses-git/lm/query lm.wmt11.kenlm5.de \
null
Loading statistics:
user    0
sys     1.00059
rss  6142736 kB
,=1613017 1 -1.77036    und=2126685 2 -1.89565  den=2039297 3 -2.11198
haben=2505264 4 -2.60478        .=434224 3 nan  Total: nan OOV: 0
After queries:
user    0
sys     1.00059
rss  6142780 kB
Total time including destruction:
user    0
sys     1.00092
rss     1480 kB


@Ken, can you have a look at this, and do you need more information (or
the model) to pinpoint the problem?


best wishes,
Rico

_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to