[Moses-support] mkcls - usable for bilingual class creation?
Hi. The original Och paper (at least the one not in, IIRC, German) that's referenced for mkcls (at http://www.fjoch.com/mkcls.html) includes a description of using a starting alignment to improve the word classification. Is this indeed doable with mkcls, and if so, how? Thanks, -Allen (Allen W. Smith, Ph.D.) ___ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support
Re: [Moses-support] Eliminating load times for MOSES phrase table, language model and reordering model
Try Compact Table? http://www.statmt.org/moses/?n=Advanced.RuleTables#ntoc3 2017年4月5日 下午7:33,"RR"寫道: > Hi, > > I am currently working on a system to explore MOSES decoding parameter > space - running MOSES in decoding mode with one parameter set, getting a > BLEU score and rerunning with different parameters. > > More than 50% of my time is being spent repeatedly "loading" the phrase > table, language model and reordering model. For example, > > $ ./example.fr | .../moses -f ./run11.moses.ini -threads 31 > > ./trans_baseline > > gives: "Start loading text phrase table. Moses format : [37.296] seconds." > > 37 seconds just to load a 590MB phrase table, even if it is on RAMDISK. > htop indicates that for those 37 seconds, a single core is 100% utilized. > > My understanding is that some kind of optimized data structure (a hash > table of some kind?) is being created in those 37 seconds, which is then > lost and recreated when I re-run MOSES with a different value of a decoding > parameter. > > If I want to eliminate this loading time, what is the best way forward? Is > there a phrase table format that avoids a significant loading time (i.e. a > 500MB phrase table loads in < 1 sec)? Should I try to find a way to run a > single MOSES server but with different decoding parameters? I have a system > with 130GB RAM and I am using a small phrase table that has been filtered > for my decoding set. > > Thanks in advance, > > Roko > > ___ > Moses-support mailing list > Moses-support@mit.edu > http://mailman.mit.edu/mailman/listinfo/moses-support > > ___ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support
[Moses-support] Eliminating load times for MOSES phrase table, language model and reordering model
Hi, I am currently working on a system to explore MOSES decoding parameter space - running MOSES in decoding mode with one parameter set, getting a BLEU score and rerunning with different parameters. More than 50% of my time is being spent repeatedly "loading" the phrase table, language model and reordering model. For example, $ ./example.fr | .../moses -f ./run11.moses.ini -threads 31 > ./trans_baseline gives: "Start loading text phrase table. Moses format : [37.296] seconds." 37 seconds just to load a 590MB phrase table, even if it is on RAMDISK. htop indicates that for those 37 seconds, a single core is 100% utilized. My understanding is that some kind of optimized data structure (a hash table of some kind?) is being created in those 37 seconds, which is then lost and recreated when I re-run MOSES with a different value of a decoding parameter. If I want to eliminate this loading time, what is the best way forward? Is there a phrase table format that avoids a significant loading time (i.e. a 500MB phrase table loads in < 1 sec)? Should I try to find a way to run a single MOSES server but with different decoding parameters? I have a system with 130GB RAM and I am using a small phrase table that has been filtered for my decoding set. Thanks in advance, Roko ___ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support
Re: [Moses-support] Support Moses and GPU on cloud
Dear Marcin and Moses community, Thanks for the tips! Yeah, g2.8xlarge is painfully expensive... Training on separate instances sounds more reasonable. Now, I've to explain to the devs why I need 2 instances ;P Regards, Liling On Tue, Apr 4, 2017 at 4:25 PM, liling tanwrote: > Dear Marcin and Moses community, > > Are you running on g2.8xlarge on AWS? > > I think I went to the cheap g2.2xlarge and 15GB RAM is a little too low > for MGIZA++ , taking forever... I think I've got to recreate a new larger > instance. > > Regards, > Liling > > > > > ___ Moses-support mailing list Moses-support@mit.edu http://mailman.mit.edu/mailman/listinfo/moses-support