Hi Maria

* Looking for MT/NLP opportunities *
Hieu Hoang
http://moses-smt.org/


On 21 March 2017 at 14:23, Maria Braga <[email protected]> wrote:

> Hi Hieu,
>
> I am Maria and work at Unbabel. We've met in Prague at the MT Marathon
> last year.
>
> I am currently working with the feature weight-overwrite to switch between
> domain-adapt models. Attached the moses.ini I am using.
>
> The weights of the models are zero in moses.ini and the real weights
> (obtained during tuning) are sent along with the request to the
> mosesserver. E.g.:
>
> params = {"text": '<weight-overwrite weights="Distortion0=
> 0.0511441548642292 LR0= 0.0437515592800108 0.0279881955302978
> 0.0382101724410126 0.0867390742233579 0.00941393194304905
> 0.0112830102073967 LR2= 0.0618355738339385 0.000426160492546368
> 0.0247603300469584 0.0582980822135267 -0.00453377850874854
> 0.0241469832817032 LM0= 0.0315944725599535 LM2= 0.0487458450881901
> WordPenalty0= -0.16828130062377 PhrasePenalty0= 0.0212747640668878 TM0=
> 0.0334756678269972 0.052272015266689 0.0504694581232027 0.00714682849923318
> TM2= 0.0206656704919728 0.0496986821183746 0.0664314672807498
> 0.00741282118720342 UnknownWordPenalty0= 1"/>Hello world'}
>
> When I increase the number of models (in moses.ini) loaded, I observe the
> decoding time increasing. Example:
>
>    - With a general model and two domain adapt model loaded, the decoding
>    time was 1.95s
>    - When added a third domain adapt model, decoding time went up to 2.80s
>
> Some notes: This values are the average decoding time for the same test
> corpus with ~650 sentences. The second bullet was a test performed with the
> three domain adapt models being the same (loaded three times the same
> model) and being different (two were the same model and a different model).
> Also the machine I am running this experiments have enough memory (the
> loaded models use a bit more than half of the available memory and there is
> enough CPUs).
>
> The questions are:
>
>
>    1. Do you know this "weight-overwrite" feature?
>
> yes, this over-rides the weights in the ini file. I don't know much about
the moses server though.

>
>    1. Why does by increasing the number of models from two to three
>    (initialised at zeros) the decoding time increases?
>
> Even though the weights are zero, the models are still used. Phrase-pairs
need to be lookup up, new translations are created from the phrase-pairs
etc.

>
>    1. If you don't know, can you point where should I look to find the
>    answers to these questions?
>
> Best if you email the moses mailing list (cc'ed) so other people can chip
in too

>
> Cheers,
>
> Maria Braga
>
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to