Hi all,

I've recently posted a query on the problem I was having with MERT. I found
a similar thread on the list (by Cyrine Nasr) and posted the question on the
same thread.

Not to be repeating myself, but the problem I was facing was: MERT step was
crashing after running exactly one iteration. The error message I got was
something like this:

Executing: 
/home/pkolachi/Documents/softwares/tools/moses-multithreaded/trunk/mert/mert
-d 14  --scconfig case:true -n 20 --ffile run1.features.dat --scfile
run1.scores.dat --ifile run1.init.opt > mert.out 2> mert.log
 Exit code: 3
Failed to run mert at /home/svenkata/tools/mosesdecoder/trunk/scripts/
scripts-20110317-1219/training/mert-moses.pl line 726.

I installed a recent version of decoder with multi-threading after which
this problem popped up.
I am really unsure of the revision numbers that I used, so that I will just
say that:
 1. The latest version (build 4251 .after the changes by Heafield) was used
with threading enabled. This is the version that I encounter the problem in
MERT with.

 2. The version  of Moses that I used previously is unknown to me, but I can
say that it was much before the improvements were made to MERT (Philipp
Koehn, August 11th). It was built for single thread use.

After comparing the code in both the versions of MERT, I noticed the
following changes and a slight 'anomaly'. Well, I actually need one of you
guys to verify and confirm so that I can call it a bug!

In file mert.cpp;
Previous version: read starting points at beginning of each iteration and
store it in vector start.
Current version: read starting points at beginning of each iteration and
store in in vector start
 AND also store values at beginning of 1st iteration in min and
maxvectors.  [lines 250-267]

However, each time we would want to save the min and max vectors, one would
have to *reset the pointer* to the file containing the starting points of
the iteration since it is already pointing to the last value after reading
vector start.
The two seek operations are missing, as a result of which the feature
extraction and scoring happen in iteration 1 but mert itself crashes.

I've managed to get my local copy of mert to work after making these two
changes. It would be very helpful and reassuring to me if any one could take
a look at this and say that this is the actual fix required for the problem.


Thank you.
- Prasanth

-- 
"Theories have four stages of acceptance. i) this is worthless nonsense; ii)
this is an interesting, but perverse, point of view, iii) this is true, but
quite unimportant; iv) I always said so."

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

Reply via email to