The EMS has been updated since then to be able to train using IRSTLM. I'll
change the webpage now

When using the EMS, set lm-training to
   lm-training = "$moses-script-dir/generic/trainlm-irst2.perl -cores
$cores -irst-dir $irstlm-dir -temp-dir $working-dir/tmp"

On 9 March 2013 15:11, Per Tunedal <[email protected]> wrote:

> Hi,
> I'm using IRSTML, so the EMS isn't an option, is it?
> http://www.statmt.org/moses/?n=Moses.Baseline :
> "The only wrinkle is that EMS currently (April 2012) does not support
> training with IRSTLM, so either you'll have to install and use SRILM
> (and refer to the main [EMS -> FactoredTraining.EMS]] documentation to
> see how to set up training) or see above for instructions on how to
> train the LM with IRSTLM (if you haven't already done so). "
>
> I've counted to 2 million sentences in the corpus. My computer has 8 GB
> RAM and 4 cores. I'm happy I've reverted to building a phrase-based
> model, instead of a hierarchical one.
>
> Yours,
> Per Tunedal
>
> On Sat, Mar 9, 2013, at 15:53, Hieu Hoang wrote:
> > How much memory and how many cores does your computer have? How many
> > sentences are in your parallel corpus?
> >
> > The training shouldn't be overwhelming your computer, it should take the
> > same amount of memory & cpu whatever the data size or model. However, the
> > hierarchical model takes about 10 time longer to train than the
> > phrase-based, and the models are 10 times bigger.
> >
> > when tuning, imo, you need at least 4gb to run the hierarchical decoder,
> > but much more memory is better. With phrase-based, you can easily do it
> > with 2GB.
> >
> > if you're trying to create an online model, to run fast, you'll need a
> > computer that has enough memory as the size of the model. This could be
> > 100GB or 200GB or more for hiero models.
> >
> > Another way is to use suffix arrays, at the cost of slower overall
> > decoding
> > time.
> >
> > In summary, hiero models are a challenge to run. Phrase-based is easy.
> >
> >
> >
> > ps. if you're starting out with Moses, I recommend using the EMS.
> >    http://www.statmt.org/moses/?n=FactoredTraining.EMS
> >  you should take an example file and change it to your needs. The files
> >  are
> > in the moses repository under
> >    scripts/ems/example
> > The EMS saves you from have to learn all the arguments in every script.
> > However, it calls those scripts so you can see exactly how to call them
> > if
> > you want.
> >
> >
> >
> > On 9 March 2013 13:30, Per Tunedal <[email protected]> wrote:
> >
> > >
> > > Hi,
> > > the training seems to be an overwhelming task for my computer. If it
> > > ever succeeds, I will have to undertake the even more demanding task of
> > > tuning. Can anything be done to accelerate it?
> > >
> > > Specifically, I wonder if it's feasible to prune the translation table
> > > before doing the tuning.
> > >
> > > Yours,
> > > Per Tunedal
> > >
> > > PS I've abandoned the idea of building a Hierarchical phrase model, I'm
> > > now trying to make a phrase-based system. I suppose that would use less
> > > resources.
> > >
> > > _______________________________________________
> > > Moses-support mailing list
> > > [email protected]
> > > http://mailman.mit.edu/mailman/listinfo/moses-support
> > >
> >
> >
> >
> > --
> > Hieu Hoang
> > Research Associate
> > University of Edinburgh
> > http://www.hoang.co.uk
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
>



-- 
Hieu Hoang
Research Associate
University of Edinburgh
http://www.hoang.co.uk
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to