Hi Ondrej I successfully completed a mert run using your script. Looks to be working mostly ok.
The only hiccup was that in an earlier run I think moses crashed and left an empty features.list. Next time around it didn't bother to do the 'show- weights' run, and used the empty features.list, which caused tuning to crash after the first iteration. Is there a need to have a facility for reusing features.list? If so, then there needs to be checks to ensure we don't use an incorrect/broken one from earlier. In order to get better feedback, I'd suggest getting your new version into the trunk as soon as possible. As long as you're monitoring the moses list to answer support questions, then it should be fine to make mert-moses-ondrej the only mert script. If all else fails, we can always revert to the previous version, cheers - and thanks for fixing mert! Barry On Sunday 03 Jul 2011 22:18:22 Ondrej Bojar wrote: > Hi, all (esp. Barry and Nicola), > > at last, I've managed to finish my little cleanup of mert-moses.pl. The > cleaned version is available under the name mert-moses-ondrej.pl for a > while, to avoid breaking other people's pipelines. > > Please test mert-moses-ondrej.pl and report problems. > > The main difference is that I now rely on the flag -show-weights of > moses instead of parsing moses.ini and mangling with 'triples'. This > should be much less fragile in the long term and one more benefit is > that moses actually checks all the models before mert starts. (Binarized > models are not checked as thoroughly as they probably should, but that's > further deep in moses.) > A possible drawback is that the node where mert runs may not be strong > enough for this test in parallel environments. > > I've also added the support for manually setting bounds of weights in > cmert random restarts, this works with both mert-moses.pl and > mert-moses-ondrej.pl > > Cheers, > O. > _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
