Hi, I just checked into svn support for binarizing hierarchical rule tables into filter-model-given-input.perl and experiment.perl.
For filter-model.perl, you may use the switch: -Binarizer "/home/pkoehn/moses/CreateOnDisk/src/CreateOnDiskPt 1 1 5 100 2" For mert.perl, you may use the switch --filtercmd '/home/pkoehn/moses/scripts/training/filter-model-given-in put.pl -Binarizer "/home/pkoehn/moses/CreateOnDisk/src/CreateOnDiskPt 1 1 5 100 2" For experiment.perl, you may use the setting ttable-binarizer = "$moses-src-dir/CreateOnDisk/src/CreateOnDiskPt 1 1 5 100 2" -phi On Fri, May 28, 2010 at 1:44 PM, Hieu Hoang <[email protected]> wrote: > the glue grammar is very small so doesn't need to be binarised. The > 'real' translation model should be binarised with > > CreatOnDiskPt > > to minimise RAM. My ini file has this as the ttable-file section: > > [ttable-file] > 2 0 0 2 ..*path*../pt.hiero > 6 0 0 1 ..*path*../glue.hiero > > > On 28/05/2010 02:30, Hwidong Na wrote: >> Thanks for your reply. >> >> I've tried the hierarchical model using larger corpura (about 2M pairs >> of sentences) as described in the website and CreatOnDiskPt works fine. >> However, moses_chart during mert-moses.perl gives an error: >> >> moses_chart: OnDiskWrapper.cpp:48: bool >> OnDiskPt::OnDiskWrapper::OpenForLoad(const std::string&): Assertion >> `m_fileSource.is_open()' failed. >> >> when the file path is for the glue grammar. >> >> Should I also create the on-disk glue grammar when I use an on-disk >> phrase table? >> >> By the way, the flag TRAINING:binarize-all seems to make it possible >> that running the hierarchcal models using the experiment.perl. >> >> Best regards, >> > _______________________________________________ > 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
