Hi,
the current version of filtering in EMS operates on the phrase table
and reordering table, and not the moses.ini configuration file.
This has the advantage that the same filtering step can be re-used
with different language models or differently tuned model weights.
But since the filtering script expects a configuration, such a dummy
configuration file is created, which points to a dummy language model
file --- in fact the current working directory as you noticed.
I have not encountered any problems with this setup, but if you notice
problems with using a Windows server file system, I suggest to use
a different dummy language model file. Please change line 2509
in experiment.perl.
Maybe the following works:
$cmd .= "-lm 0:3:$dir/steps/$VERSION/config.$VERSION "; # dummy
-phi
On Wed, Nov 7, 2012 at 7:23 AM, Guchun Zhang <[email protected]> wrote:
> Hi,
>
> I'd like to report a trivial problem at the filter stage of the latest Moses
> version.
>
> At the filter stage, train-model.perl will use the working directory as the
> temporary language model (which I don't understand) if it exists and is
> non-zero. Line 430 of train-model.perl will do the checking and report
> error. When the working directory is on a network drive (not NFS and in my
> case a windows server), the size of this directory and all sub-directories
> is zero, which is the root of this problem. Pointing the lm input to a lm
> file under lm directory or moving the working directory onto a local or NFS
> location will solve this problem. Storing models on a network drive is
> favourable when local disks run out of space.
>
> Cheers,
>
> Guchun
>
> _______________________________________________
> Moses-support mailing list
> [email protected]
> http://mailman.mit.edu/mailman/listinfo/moses-support
>
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
