Hi Folks

Just to clarify...

mert-moses-multi.pl is for use with the interpolated scorer implemented by 
Christophe Servan, and described in his MTM paper last year. It doesn't have 
anything to do with multi-threading. If you're not using the interpolated 
scorer, just use mert-moses.pl.

You can pass the --threads=NUMBER argument to mert-moses.pl and it will use 
multi-threaded mert (i.e. the inner loop) which will run the random restarts 
in paralllel (threads are compiled in by default now). This can be useful if 
you want lots of random restarts, or have a large tuning set or large nbest 
lists. However,  for most setups, it's decoding that's the bottleneck in 
tuning.

To pass parameters to the decoder in tuning when using experiment.perl, you 
can add a line like the following

decoder-settings = "-threads 4"

to the TUNING stanza.

And yes, the documentation on tuning is sadly out-of-date.

cheers - Barry



On Thursday 05 Jul 2012 20:01:45 Alexander Fraser wrote:
> I think there is a version of MERT that is multi-threaded? I am not
> sure that it supports batch MIRA (see recent message from Colin
> Cherry) - this is good and fast (and free :-) ) and perhaps doesn't
> need to be parallelized?
> 
> BTW, has anyone figured out how to pass the threads parameter through
> to the decoder in tuning when using experiment.perl? I hardcoded it to
> use $__THREADS in mert-moses.pl in the decoder command line, line
> 1069.
> 
> Cheers, Alex
> 
> 
> On Thu, Jul 5, 2012 at 1:50 PM, Tom Hoar
> 
> <[email protected]> wrote:
> >  mert-moses.pl and mert-moses-multi.pl both have the command argument
> >
> >  --threads=NUMBER          ... Use multi-threaded mert (must be compiled
> >  in)
> >
> >  Can someone explain the caveat "must be compiled in"?
> >
> >  Also, since both mert-moses.pl and mert-moses-multi.pl have the
> >  --threads argument, what is the difference between the two and is there
> >  a situation when one is preferred over the other?
> >
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to