Hi, the standard pipeline filters the phrase translation table for tuning and evaluation.
The result of tuning, however, is just the feature weights. None of the model files should be changed. Specifically, you should not just use the moses.ini file in the tuning directory, since this is uses a phrase table filtered for the tuning set. In the standard experiment.perl pipeline, the script $MOSES/scripts/ems/support/substitute-weights.perl substitutes weights in the original moses.ini with the tuned weights. -phi On Tue, Nov 17, 2015 at 2:17 AM, Apurva Joshi <[email protected]> wrote: > hello , > > initially i trained with corpora containing 7 lac sentnses...i trained > it..and its giving good translation... > > then for tuning i took new small size corpora containing 2k lines .... > > now i run tuning with this new processed corpora as input along with my > moses.ini (generated after training) as input ....then mert-work folder is > generated with run6.moses.ini .... > > along with this ,phrase table.gz and reordering table is also generated in > mert-work folder... > > now problem is .... > > 1st problem : > > before tuning , size of phrase table is 2.5Gb > > after tuning , size of phrase table is 256.5Mb > > 2nd problem: > > now while running moses, should i use moses.ini(in mert-work/filtered > folder) OR moses.ini(run6.moses.ini) ...? > > i ran moses using both above mentioned moses.ini.... > from both m not getting proper o/p... > eg > > ip = hello > > op = > > *BEST TRANSLATION: hello|UNK|UNK|UNK * > before tuning , it was giving me correct translation of hello i.e > > op= नमस्ते (hindi translation of hello).... > > > so whether is it taking memory problem or what ?? plzz help... > > > > > _______________________________________________ > 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
