Moses-support Digest, Vol 104, Issue 12

Send Moses-support mailing list submissions to
moses-support@mit.edu

To subscribe or unsubscribe via the World Wide Web, visit
http://mailman.mit.edu/mailman/listinfo/moses-support
or, via email, send a message with subject or body 'help' to
moses-support-request@mit.edu

You can reach the person managing the list at
moses-support-owner@mit.edu

When replying, please edit your Subject line so it is more specific
than "Re: Contents of Moses-support digest..."


Today's Topics:

1. Re: Moses infinite loop (Hieu Hoang)
2. BLEU score is lower than normal for baseline system (FR to
EN) (Davood Mohammadifar)


----------------------------------------------------------------------

Message: 1
Date: Tue, 09 Jun 2015 21:28:48 +0400
From: Hieu Hoang <hieuhoang@gmail.com>
Subject: Re: [Moses-support] Moses infinite loop
To: "Bc. Robert ADAMEC" <adamecr@students.zcu.cz>, moses-support
<moses-support@mit.edu>
Message-ID: <55772250.6020309@gmail.com>
Content-Type: text/plain; charset=utf-8; format=flowed

ah, you're decomposing the translation into factors. You have to be
careful that the number of translation rules it create doesn't explode.
Try these things
1. Set table-limit for each phrase-table to a small number, eg.
PhraseDictionaryCompact .... table-limit=5
2. Prune the generation table so that it keeps only the best 5
output factor for each input factor.
Once you know that your experiments can run in the time and memory you
need, you can increase these limits

On 09/06/2015 20:24, Bc. Robert ADAMEC wrote:
> Hello,
>
> this is the exact command:
>
> /opt/moses-3.0/scripts/training/mert-moses.pl /root/faktor_6/corpus/heldout.src /root/faktor_6/corpus/heldout.ref \
> /opt/moses-3.0/bin/moses /root/faktor_6/experiment/model/moses.ini \
> --working-dir /root/faktor_6/tmp \
> --decoder-flags "-threads 6 -v 100" \
> --rootdir /opt/moses-3.0/scripts -mertdir /opt/moses-3.0/bin \
> --maximum-iterations 10
>
> The moses.ini file and the log file are attached.
>
> It seems that the tool maybe is not stuck but proceeds so slowly that it is useless anyhow, strange thing is that it occurs on a certain point....until the hypothesis on the row 4844 it runs fast and then it slows heavily, the remaining hypotheses seen in the log from that row have been processed in about 16 hours while the hypotheses before that row were processed in like 5 minutes, maybe even less.
>
> Still do not get why is that happening and it frustrates me because it is part of my diploma thesis.
>
> Best regards,
>
> Robert Adamec
>
>
>
> ?ter?, 9 ?erven, 2015 08:27 CEST, Hieu Hoang <hieuhoang@gmail.com> napsal:
>
>> so u binarised the phrase table and re-ordering model. Did you change the
>> moses.ini file to tell the decoder to use the binarised models?
>>
>> to avoid confusion, please send me the exact command you used and the exact
>> moses.ini file you used
>>
>> Hieu Hoang
>> Researcher
>> New York University, Abu Dhabi
>> http://www.hoang.co.uk/hieu
>>
>> On 9 June 2015 at 10:06, Bc. Robert ADAMEC <adamecr@students.zcu.cz> wrote:
>>
>>> Hello,
>>>
>>> the execution command is described in the first e-mail that I have send
>>> you. (the first e-mail in this e-mail chain)
>>>
>>> I have further observed the behavior of the tool and it seems to move
>>> through the hypothesis, but at very very slow rate. It runs normally for a
>>> while, then it hits the hypothesis, which can be seen in the log from
>>> previous e-mails, and starts to run very slowly, like 3 hypothesis per 10
>>> hours. I do not know if it would ever complete the decoding stage in the
>>> for cycle that I have posted previously but nevertheless I do not have time
>>> to let it run for weeks if it moves somewhere.
>>>
>>> By debug messages you mean the messages added by me or normal debug
>>> messages which Moses outputs with -v 3? You can see the normal debug output
>>> in the log from the first e-mail.
>>>
>>> Best regards and thank you for your time,
>>>
>>> Robert Adamec
>>>
>>>
>>> On Monday, June 8, 2015 18:27 CEST, Hieu Hoang <hieuhoang@gmail.com>
>>> wrote:
>>>
>>>> What is the exact command you're executing? Can u please send me the
>>> debug
>>>> messages that Moses output
>>>> On 8 Jun 2015 19:54, "Bc. Robert ADAMEC" <adamecr@students.zcu.cz>
>>> wrote:
>>>>> Hello,
>>>>>
>>>>> I made the compact versions of the phrase tables and reordering models,
>>>>> the language models were already binarised. However it did not help,
>>> the
>>>>> tool still "freezes" on the same spot and stays there.
>>>>>
>>>>> I have let it run for almost 24 hours and it moved nowhere. I really do
>>>>> not think that it would change with time, the experiments that I used
>>>>> earlier run about 24 hours total, for this experiment I added factored
>>>>> corpus for both sides, previously it was only for the source language,
>>> and
>>>>> I do not think that it should impact the performance so heavily.
>>>>>
>>>>> I have tried to put some debug messages into moses itself, namely into
>>> the
>>>>> DecodeStepGeneration class around the lines 149 - 170 (the for (size_t
>>>>> currIter = 0 ; currIter < numIteration ; currIter++) cycle) and I
>>> found out
>>>>> that the tool freezes because it gets stuck in this cycle forever
>>> after a
>>>>> while (for the hypothesis that you can see as last in the log from the
>>>>> previous e-mail) and never leaves it.
>>>>>
>>>>> I do not know what is the reason and cannot find it out.
>>>>>
>>>>> Any ideas?
>>>>>
>>>>> Best regards,
>>>>>
>>>>> Robert Adamec
>>>>>
>>>>>
>>>>>
>>>>> ?ter?, 2 ?erven, 2015 14:05 CEST, Hieu Hoang <hieuhoang@gmail.com>
>>>>> napsal:
>>>>>
>>>>>> please subscribe to the mailing list before posting to it. You can
>>>>>> subscribe here:
>>>>>> http://mailman.mit.edu/mailman/listinfo/moses-support
>>>>>>
>>>>>> To answer your question - how do you know it is in an inifinite loop,
>>>>>> rather than just taking a very long time? From the moses.ini file, it
>>>>>> seems the phrase tables are not binarized. This will cause loading to
>>>>>> take a very long time and use a lot of memory. You should binarize
>>> your
>>>>>> phrase-table, reordering model and LM.
>>>>>>
>>>>>> Also, if you are new to moses, please work through the tutorial here
>>> to
>>>>>> get a better understanding of what moses is doing
>>>>>> http://www.statmt.org/moses/?n=Moses.Baseline
>>>>>>
>>>>>>
>>>>>> On 29/05/2015 14:35, Bc. Robert ADAMEC wrote:
>>>>>>> Hello,
>>>>>>>
>>>>>>> I have encountered some trouble while working with Moses 3.0
>>>>> optimization.
>>>>>>> When I use the configuration which you can see in the attached
>>>>> moses.ini the Moses tool runs infinitely long in the run 1 and is
>>> stuck on
>>>>> the Loading table into memory line. Do you know what could be the
>>> problem?
>>>>> For the "-v 100" log see the atached log.txt
>>>>>>> When I tweak some of the config options in the moses.ini file I get
>>>>> segfault, for example if I change the GenerationModel from current 1
>>> -> 0
>>>>> to 0,1 -> 0.
>>>>>>> The source corpus has the "word|factor" format for both used
>>> languages
>>>>> so there should be no problems, the factor is an int number.
>>>>>>> Can you imagine any reason why it could be behaving this way and
>>> what
>>>>> could have gone wrong?
>>>>>>> I launch the optimization via this command:
>>>>>>>
>>>>>>> /opt/moses-3.0/scripts/training/mert-moses.pl
>>>>> /root/faktor_6/corpus/heldout.src /root/faktor_6/corpus/heldout.ref \
>>>>>>> /opt/moses-3.0/bin/moses
>>> /root/faktor_6/experiment/model/moses.ini
>>>>> \
>>>>>>> --working-dir /root/faktor_6/tmp \
>>>>>>> --decoder-flags "-threads 6 -v 0" \
>>>>>>> --rootdir /opt/moses-3.0/scripts -mertdir /opt/moses-3.0/bin \
>>>>>>> --maximum-iterations 10
>>>>>>>
>>>>>>> Best regards,
>>>>>>>
>>>>>>> Robert Adamec
>>>>>> --
>>>>>> Hieu Hoang
>>>>>> Researcher
>>>>>> New York University, Abu Dhabi
>>>>>> http://www.hoang.co.uk/hieu
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>
>>>
>>>
>>>
>>>
>>>
>
>
>
>

--
Hieu Hoang
Researcher
New York University, Abu Dhabi
http://www.hoang.co.uk/hieu



------------------------------

Message: 2
Date: Wed, 10 Jun 2015 12:14:47 +0000
From: Davood Mohammadifar <davood_mf@hotmail.com>
Subject: [Moses-support] BLEU score is lower than normal for baseline
system (FR to EN)
To: "moses-support@mit.edu" <moses-support@mit.edu>
Message-ID: <SNT150-W814D98A89615A878DC2EF08CBD0@phx.gbl>
Content-Type: text/plain; charset="iso-8859-1"

Hello everyone

I executed all instructions for building baseline system (French to English SMT based on moses manual) with following system:

CPU: core i5 2400
RAM: 4GB
OS: Ubuntu 14.04 32-bit
Moses: v3.0 (I used it in virtual machine that statmt.org released)

After executing all of commands (training, tuning and testing) , the BLEU score was 19.78. In moses manual, it is said that score 23.5 (with 2 points difference) is ok.
So my score is lower than normal.

Why my BLEU score is lower than normal? Is it related to virtual machine or my system?


Thank you

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20150610/f5a94cf4/attachment-0001.htm

------------------------------

_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support


End of Moses-support Digest, Vol 104, Issue 12
**********************************************

Tweet