Hi all, I’m chiming in a bit late on this conversation, but thought what I have could help.
At Autodesk we modified the training process a few years back to avoid the necessity of using so much temp disk space and to reduce the overall disk I/O during training—all with the premise that in a stable environment we don’t care about the intermediary files. Our code is available on GitHub at https://github.com/venyz/ADSKMosesTraining Have a look specifically at https://github.com/venyz/ADSKMosesTraining/blob/master/scripts/training/train-model.perl around lines 500–600, where we launch a bunch of processing steps in parallel setting up the data transfer between tools via unix named pipes. For this to work properly, we also had to modify some of the other training workflow tools, but never had the time and resources to reintegrate our changes in the Moses master (at the time Moses wasn’t even on Git yet). Though I doubt you can use this directly with the current moses training tools, I hope this gives you inspiration and helps solve Alexander’s disk space issue. Also, Tom you said that you’re working on updating train-model.perl anyway, so probably our modifications could be something you can consider using. For more information on what we actually did and a numerical comparison of the disk usage statistics with and without our modifications, check Section 2.1 in Zhechev, Ventsislav. 2012. Machine Translation Infrastructure and Post-editing Performance at Autodesk. In Proceedings of the AMTA 2012 Workshop on Post-editing Technology and Practice (WPTP ’12), eds. Sharon O’Brien, Michel Simard and Lucia Specia. San Diego, CA. (http://amta2012.amtaweb.org/AMTA2012Files/html/5/5_paper.pdf) Cheers, Ventzi ––––––– Dr. Ventsislav Zhechev Computational Linguist, Certified ScrumMaster® Platform Architecture and Technologies Localisation Services MAIN +41 32 723 91 22 FAX +41 32 723 93 99 http://VentsislavZhechev.eu Autodesk, Inc. Rue de Puits-Godet 6 2000 Neuchâtel, Switzerland www.autodesk.com > 27.02.2015 г., в 12:49, [email protected] написал(а): > > Message: 3 > Date: Fri, 27 Feb 2015 08:25:01 +0700 > From: Tom Hoar <[email protected]> > Subject: Re: [Moses-support] My phrase-table.tgz is 20-bytes long > To: [email protected] > Message-ID: <[email protected]> > Content-Type: text/plain; charset="utf-8" > > The temp space challenge with train-model.perl is not only how much temp > space is needed, but also where are the temp files created? > train-model.perl seems unpredictable about where temp files are placed. > > In general, the /tmp folder goes unused unless you specify it with the > "--temp-dir" option. Even when specified, only the sort functions use > the configuration it. Without setting the option,e sort temp files go to > the "--model-dir" folder, which by default is "--root-dir / model". > Finally, if not specified, --root-dir defaults to the current working > directory (i.e. cwd or "."). > > That said, the temp folder for extraction step 3 is different. Its temp > files default to a newly-created "tmp" folder under the parent directory > of the path value set with "--extract-file". If "--extract-file" is not > set, the default is "--model-dir / extract", which also follows the > "--model-dir" path defined above. > > Because of this complexity, we advise our customers to use hardware > configuration that relies on one large root folder partition... i.e. 1 > or 2 TB mounted as "/" root without any other mounts. This gives Moses > one contiguous pool of temp storage independent of the configuration > complexities. I acknowledge this recommendation violates normal system > administrators' configurations, but without significant changes to > train-model.perl, this is our best recommendation. > > > > On 02/27/2015 04:28 AM, Barry Haddow wrote: >> Hi Alexander >> >> From the error logs, it looks as though alignment went fine, the >> training pipeline reports 24860460 lines of aligned bitext. Since the >> extract files were empty, I'd suggest that extraction crashed, and the >> most likely is that it ran out of disk. I'm not sure what happened to >> the error messages. >> >> For 25M sentence pairs, the final phrase table could easily be 30G and >> the intermediate files are larger. You probably need more like 500G to >> be safe. >> >> I would follow Tom's advice and start with a much smaller corpus to >> see how the process works. Also, for the full corpus, you could look >> in to fast_align (https://github.com/clab/fast_align) for alignment as >> it is much faster than mgiza (e.g. 2 days versus 2 weeks), and use EMS >> for large jobs since it's much easier to restart a failed step. >> >> cheers - Barry >> >> On 26/02/15 15:06, ????????? ??????? wrote: >>> >>> Hi Barry! >>> >>> Here you can download training.out >>> https://www.dropbox.com/s/d0f0n99x4wbw3mo/training.out.gz?dl=1 >>> >>> I have about 50 Gb of free space in working dir. >>> >>> >>> 2015-02-25 17:19 GMT+07:00 Barry Haddow <[email protected] >>> <mailto:[email protected]>>: >>> >>> Hi Alexander, >>> >>> It looks like something went wrong at the extract stage. If you >>> could make your training.out available then we can look for clues. >>> >>> Could the system have run out of disk space, either in the >>> working directory or in /tmp? A lot of space is required to build >>> the extract files and phrase tables. >>> >>> cheers - Barry >>> >>> >> >> >> >> _______________________________________________ >> Moses-support mailing list >> [email protected] >> http://mailman.mit.edu/mailman/listinfo/moses-support > > -------------- next part -------------- > An HTML attachment was scrubbed... > URL: > http://mailman.mit.edu/mailman/private/moses-support/attachments/20150227/2f4f190f/attachment-0001.htm
_______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
