I've often pondered the thought of a stripped-down Moses, but stripped-down
Moses != faster, smaller Moses.

I think a stripped-down Moses would be cleaner and allow easier addition of
new functionality. (And then it won't be stripped-down anymore...)
In a similar vein, I'm refactoring the decoder to be cleaner and allow
easier additions, while not taking away many functionality. There's 5% less
code the last time I looked.

I think a faster, smaller decoder would need to be really focused on it
objective for small devices and probably requires a complete rewrite.
However, it could still reuse the training & tuning from Moses. Anyone
interested in doing it?



On 3 April 2013 12:45, Marcin Junczys-Dowmunt <[email protected]> wrote:

> Maybe it's time for a mini-moses fork. Instead of incorporating everything
> throw everything out that's not necessary to produce a translation :)
>
> W dniu 03.04.2013 13:34, Hieu Hoang pisze:
>
>> it is impressive, and it would make sense. Decoding would be simple
>> without non-local features like LMs.
>>
>> And on my tablet, it's using 30MB RAM. Moses would be hard pushed to use
>> 300MB
>>
>>
>> On 3 April 2013 12:23, Marcin Junczys-Dowmunt <[email protected]<mailto:
>> [email protected]>> wrote:
>>
>>     Still impressive in my opinion. I wonder whether they actually have an
>>     LM. Maybe they just intersected it with the phrase table and put some
>>     more weights in the table.
>>
>>     W dniu 03.04.2013 13:17, Hieu Hoang pisze:
>>     > The offline version seems to have multi-word phrases, but a bad
>>     LM, eg:
>>     >   source (vietnamese): xe hoi
>>     >   online & offline: car
>>     >
>>     >   source: brown fish
>>     >   online: poisson brun
>>     > offline: brun poissons
>>     >
>>     >   source: do you have the time
>>     >   online: avez-vous le temps
>>     > offline: *ne* vous avez le temp
>>     >
>>     > source: where is the nearest toilet please
>>     >   online: oů est le plus proche toilettes s'il vous plaît
>>     >   offline: oů est le plus proche *de* toilettes *veuillez*
>>
>>     >
>>     > Still, it's good people are finally focusing on small devices. Maybe
>>     > you'll get a call from Google...
>>     >
>>     > On 29 March 2013 17:43, Marcin Junczys-Dowmunt
>>     <[email protected] <mailto:[email protected]>
>>     > <mailto:[email protected] <mailto:[email protected]>>> wrote:
>>     >
>>     >     Something very recent on the topic:
>>     >
>>     >
>>     http://howto.cnet.com/8301-**11310_39-57576627-285/use-**
>> google-translate-offline-by-**downloading-language-packs/<http://howto.cnet.com/8301-11310_39-57576627-285/use-google-translate-offline-by-downloading-language-packs/>
>>     >
>>     >     I am impressed, about 150 MB per language pack. The
>>     translation is
>>     >     soso,
>>     >     compared to the online version, but usable.
>>     >
>>     >     W dniu 07.03.2013 08:35, Marcin Junczys-Dowmunt pisze:
>>     >     > Memory usage is usually reported to be about the same (nearly
>>     >     nothing
>>     >     > for both phrase tables), but as Kenneth mentioned there is
>>     a hidden
>>     >     > usage for both phrase tables that manifests in the disk cache.
>>     >     In theory
>>     >     > if you run moses long enough with many many sentences the
>>     old phrase
>>     >     > table may eventually end up in memory completely (provided you
>>     >     have so
>>     >     > much memory). The same is true for the compact phrase
>>     table. But
>>     >     since
>>     >     > the compact phrase table is 6-7 times smaller it will use 6-7
>>     >     times less
>>     >     > memory in an asymptotic scenario. So in theory, if the
>>     small phrase
>>     >     > tables fits in your RAM and old phrase table does not
>>     (which is
>>     >     quite
>>     >     > likely so), the compact one should be always much faster after
>>     >     some time
>>     >     > of continuous employment.
>>     >     >
>>     >     > Concerning your other question about the crash. You built
>>     the phrase
>>     >     > table on another system, right? So the problem is that in the
>>     >     trunk the
>>     >     > binary formats are not compatible between 32 and 64 bit
>>     systems.
>>     >      I have
>>     >     > versions of the phrase table and the reordering model that are
>>     >     binary
>>     >     > compatible between 32 and 64 bit systems, but I haven't
>>     put this
>>     >     into
>>     >     > the trunk. It is only available in the WIPO branch which
>>     in turn
>>     >     is a
>>     >     > lot behind the current moses development. Frankly, right now I
>>     >     do not
>>     >     > have much time available to make that work in the main moses
>>     >     > distribution, but I can take a look. However I cannot promise
>>     >     anything.
>>     >     >
>>     >     > W dniu 07.03.2013 07:38, myounggun jang pisze:
>>     >     >> We use instruments tools in xcode. we check the
>>     allocation memory.
>>     >     >> I don't know disk cached is excluded.
>>     >     >> I translate new sentences while measuring.
>>     >     >>
>>     >     >> On Thu, Mar 7, 2013 at 2:54 PM, Kenneth Heafield
>>     >     <[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>>
>>     >     >> <mailto:[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>>>**> wrote:
>>     >     >>
>>     >     >>      How did you measure memory usage?  I'm guessing it
>>     >     excludes the disk
>>     >     >>      cache that the binary phrase table functionally depends
>>     >     on.  Also,
>>     >     >>      were
>>     >     >>      you translating new sentences while measuring or
>>     >     previously translated
>>     >     >>      sentences?
>>     >     >>
>>     >     >>      On 03/07/13 00:47, myounggun jang wrote:
>>     >     >>      > Thank you for your help.
>>     >     >>      > I tested compact table on iPhone.
>>     >     >>      > however memory usage, loading time and translation
>>     time is
>>     >     >>      similar with
>>     >     >>      > binary phrase table.
>>     >     >>      >
>>     >     >>      > On Wed, Mar 6, 2013 at 7:28 PM, Marcin Junczys-Dowmunt
>>     >     >>      > <[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>>
>>     >     <mailto:[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>>>
>>     >     >>      <mailto:[email protected]
>>     <mailto:[email protected]> <mailto:[email protected]
>>     <mailto:[email protected]>>
>>     >     <mailto:[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>>>>**> wrote:
>>     >     >>      >
>>     >     >>      >     Hi,
>>     >     >>      >     take a look at the charts in
>>     >     >>      > 
>> http://www.statmt.org/mtm12/**pub/marcin-mtm2012.pdf<http://www.statmt.org/mtm12/pub/marcin-mtm2012.pdf>
>>     >     >>      >
>>     >     >>      >     If you still have questions, drop me an e-mail.
>>     >     >>      >     Best,
>>     >     >>      >     Marcin
>>     >     >>      >
>>     >     >>      >     W dniu 06.03.2013 11:20, myounggun jang pisze:
>>     >     >>      >      > Hello
>>     >     >>      >      > I found that compact phrase table is 6 to 7
>>     times
>>     >     smaller
>>     >     >>      than binary
>>     >     >>      >      > phrase table.
>>     >     >>      >      > I am interested about memory usage, loading
>>     time,
>>     >     >>      translation time.
>>     >     >>      >      > Data size is also important, however memory
>>     usage and
>>     >     >>      time is more
>>     >     >>      >      > important to me.
>>     >     >>      >      > Can I get a data about comparison between
>>     two table?
>>     >     >>      >      > Thank you.
>>     >     >>      >      >
>>     >     >>      >      >
>>     >     >>      >      > ______________________________**
>> _________________
>>     >     >>      >      > Moses-support mailing list
>>     >     >>      >      > [email protected]
>>     <mailto:[email protected]>
>>     >     <mailto:[email protected]
>>     <mailto:[email protected]>**> <mailto:[email protected]
>>     <mailto:[email protected]>
>>     >     <mailto:[email protected] <mailto:[email protected]>**
>> >>
>>     >     >>      <mailto:[email protected]
>>     <mailto:[email protected]>
>>     >     <mailto:[email protected]
>>     <mailto:[email protected]>**> <mailto:[email protected]
>>     <mailto:[email protected]>
>>     >     <mailto:[email protected] <mailto:[email protected]>**
>> >>>
>>
>>     >     >>      >      >
>>     
>> http://mailman.mit.edu/**mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>     >     >>      >
>>     >     >>      > ______________________________**_________________
>>     >     >>      >     Moses-support mailing list
>>     >     >>      > [email protected]
>>     <mailto:[email protected]> <mailto:[email protected]
>>     <mailto:[email protected]>**>
>>     >     <mailto:[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>>
>>     >     >>      <mailto:[email protected]
>>     <mailto:[email protected]>
>>     >     <mailto:[email protected]
>>     <mailto:[email protected]>**> <mailto:[email protected]
>>     <mailto:[email protected]>
>>     >     <mailto:[email protected] <mailto:[email protected]>**
>> >>>
>>
>>     >     >>      > http://mailman.mit.edu/**
>> mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>     >     >>      >
>>     >     >>      >
>>     >     >>      >
>>     >     >>      >
>>     >     >>      > ______________________________**_________________
>>     >     >>      > Moses-support mailing list
>>     >     >>      > [email protected]
>>     <mailto:[email protected]> <mailto:[email protected]
>>     <mailto:[email protected]>**>
>>     >     <mailto:[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>>
>>     >     >>      > http://mailman.mit.edu/**
>> mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>     >     >>      >
>>     >     >>  ______________________________**_________________
>>     >     >>      Moses-support mailing list
>>     >     >> [email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>
>>     >     <mailto:[email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>>
>>     >     >> 
>> http://mailman.mit.edu/**mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>     >     >>
>>     >     >>
>>     >     >>
>>     >     >>
>>     >     >> ______________________________**_________________
>>     >     >> Moses-support mailing list
>>     >     >> [email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>
>>     >     >> 
>> http://mailman.mit.edu/**mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>     >     > ______________________________**_________________
>>     >     > Moses-support mailing list
>>     >     > [email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>
>>     >     > 
>> http://mailman.mit.edu/**mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>     >
>>     >     ______________________________**_________________
>>     >     Moses-support mailing list
>>     > [email protected] <mailto:[email protected]>
>>     <mailto:[email protected] <mailto:[email protected]>**>
>>     > 
>> http://mailman.mit.edu/**mailman/listinfo/moses-support<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] <mailto:[email protected]>
>>     
>> http://mailman.mit.edu/**mailman/listinfo/moses-support<http://mailman.mit.edu/mailman/listinfo/moses-support>
>>
>>
>>
>>
>> --
>> Hieu Hoang
>> Research Associate
>> University of Edinburgh
>> http://www.hoang.co.uk/hieu
>>
>>
>
>


-- 
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

Reply via email to