tuning support factored models. Can you send me your ini file. Can you make available files so I can reproduce the problem. I need the phrase-table, language model, reordering model, input sentences and reference sentences.
Best to share it via dropbox or google drive On 30 April 2013 10:47, Anand Kumar <[email protected]> wrote: > Dear Phi > > Finally I found that the error is in "-n-best-list run1.best100.out > 100" option. But this option is must for tuning and automatically > taken by mert-moses.pl. I have also tested (before tuning) my factored > trained models which is not working with "-n-best-list" option. How > can I tune my factored model using mert? thanks in advance.. > > Executing: /home/amrita/smt/tools/moses/bin/moses -v 0 -threads all > -config filtered/moses.ini -inputtype 0 -w -0.243896 -lm 0.040658 > 0.040658 0.040658 -d 0.146338 -tm 0.048779 0.048779 0.048779 0.048779 > 0.048779 0.048779 0.048779 0.048779 0.048779 0.048779 -n-best-list > run1.best100.out 100 -input-file > /media/disk/MosesSMT/TamilSMT/Fact/mertHM/temp.eng.tune > run1.out > > > ----- Original Message ----- > From: M. Anandkumar <[email protected]> > To: Philipp Koehn <[email protected]> > Sent: Sat, 20 Apr 2013 12:38:10 +0530 (IST) > Subject: Re: Doubt Regarding Tuning Factored SMT > > Dear phi > > Thank you for your information, I have created a small model which > contains 200 (Eng-Tam)sentences for training and 20 sentences for > tuning. I have attached the trained model folder (including factored > corpora and Language model) and the Tuning Error file (Details abt the > files are given below). Pls find out the solution. I like to remind u > that I am using factored SMT for only generating Tamil word Factors > (eg: LEmma|POS|Morph) and not to generate the Surface word form. > Testing is working perfectly the problem is in Tuning only. > > HMtempmodel.tar.gz Compressed folder contains factored Trained models. > lmtemp.tar.gz Compressed folder contains factored Language model > files. > mertHM.tar.gz Compressed folder contains the models for tuning > (error folder). > nohup.out Tuning error details. > temptuneHM.sh Sh file contains the command used for tuning. > > If u need any other files or details pls contact me. thanks in advance. > > > ----- Original Message ----- > From: Philipp Koehn <[email protected]> > To: M. Anandkumar <[email protected]> > Sent: Fri, 12 Apr 2013 22:20:54 +0530 (IST) > Subject: Re: Doubt Regarding Tuning Factored SMT > > Hi, > > you should not get segmentation faults with Moses, this does clearly > indicate that something went wrong. > > This may be due to usage errors or a problem with the Moses code. In > the latter case, we would be very keen on fixing it. If you can > reproduce the error on a small model (e.g., a filtered for a single > sentence), can you send your command line and model files to > moses-support? > > -phi > > On Mon, Apr 8, 2013 at 7:21 AM, M. Anandkumar > <[email protected]> wrote: > > > > Dear phi > > > > After finishing the factored model training successfully. I have > tried to tune with mert. But I had an error like given bellow. I binarized > my phrase table as per your suggestion for some of the person's question. > Still I am getting the error like this. Details abt my system is, For > training 6.5k factored sentences and tuning with .5k factored sentences. > Additionally, I am using ubuntu VM with only 1GB RAM. Is it a problem of > memory?. Thanks. > > > > > > Segmentation fault > > Register dump: > > > > EAX: 00000000 EBX: b6673124 ECX: b6673124 EDX: 00000000 > > ESI: 082c0f60 EDI: 082c0f60 EBP: b6672eb8 ESP: b6672ea0 > > > > EIP: 0809b62d EFLAGS: 00010286 > > > > CS: 0073 DS: 007b ES: 007b FS: 0000 GS: 0033 SS: 007b > > > > Trap: 0000000e Error: 00000004 OldMask: 00000000 > > ESP/signal: b6672ea0 CR2: 00000000 > > > > FPUCW: ffff037f FPUSW: ffff0024 TAG: ffffffff > > . > > . > > . > > . > > Segmentation fault > > Exit code: 139 > > The decoder died. CONFIG WAS -w -0.243896 -lm 0.040658 0.040658 0.040658 > -d 0.146338 -tm 0.048779 0.048779 0.048779 0.048779 0.048779 0.048779 > 0.048779 0.048779 0.048779 0.048779 > -- > --Regards > > Anand Kumar M > > _______________________________________________ > Moses-support mailing list > [email protected] > http://mailman.mit.edu/mailman/listinfo/moses-support > > -- Hieu Hoang Research Associate University of Edinburgh http://www.hoang.co.uk/hieu
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
