Hi Arefeh

Are your features tuneable?

I think if you run
moses --show-weights -f path-to-ini
it should tell you

cheers - Barry

On 20/08/16 15:52, arefeh kazemi wrote:
Hi Barry

Thanks so much,
moses.ini seems ok, but my feature isn't in *.best files.
These are the lines in my moses.ini:

[feature]
Aref trainFilePath=/files/entr/entr-dd-surf-tune.csv decodeFilePath=/files/entr/entr_Tune_Decoding_WithIndex.txt
UnknownWordPenalty
....
# dense weights for feature functions
[weight-file]
/old-scratch/akazemi/experiments/entr-mymodel/dd-sur/sparse/weights.txt

This is the command that I use for tuning:
nohup nice $SCRIPTS_ROOTDIR/training/mert-moses.pl <http://mert-moses.pl> entr-tune.en entr-tune.tr <http://entr-tune.tr> \
$MOSESCHART $WorkingDir/moses.ini --mertdir $MERTDIR \
--rootdir $SCRIPTS_ROOTDIR \
--batch-mira --return-best-dev --decoder-flags '-threads all -v 2' > mert.out

This is a line from debug output, red ones are my features:
BEST TRANSLATION: 639354 S -> S </s> :0-0 : c=-0.024 core=(0.000,-1.000,1.000,0.000,0.000,0.000,0.000,0.000,0.000) sourcePhrase=X </s> [0..28] 638176 [total=-5.857] core=(0.000,-33.000,41.000,-34.703,-48.214,-30.033,-39.420,16.998,-96.448) nnnnm=-3.414 prepprepm=-9.457 unalignedPenalty=79.000 cohesionPenalty=61.000 amodprepm=-10.328 detpreps=-3.297 amodamodm=-7.082 detnnm=-7.316
Best Hypothesis Generation Time: : [401.400] seconds
WRITING 100 TRANSLATION ALTERNATIVES TO run1.best100.out
N-Best Hypotheses Generation Time: : [401.645] seconds
Sentence Decoding Time: : [401.650] seconds
Translation took 1907.342 seconds


Should I check anything else?

Regards
Arefeh


On Sat, Aug 20, 2016 at 4:53 PM, Barry Haddow <[email protected] <mailto:[email protected]>> wrote:

    Hi Arefeh

    Attached.

    If you look at the files produced in your tuning run, you should
    see the following produced for each iteration:

    run2.best100.out.gz
    run2.dense
    run2.extract.err
    run2.extract.out
    run2.features.dat
    run2.init.opt
    run2.mert.log
    run2.mert.out
    run2.mira.out
    run2.moses.ini
    run2.out
    run2.scores.dat
    run2.sparse-weights
    run2.weights.txt


    (run1 may be different)

    In run2.best100.out.gz you should see the sparse feature values.
    In run2.mert.out you should see the sparse feature weights.


    cheers - Barry


    On 20/08/16 12:47, arefeh kazemi wrote:
    Hi Barry

    Thanks,
    I tried verbose and I can see my sparse feature and it's scores
    for each sentence, but the weights.txt file is still empty.
    (I use kbmira for tuning).
    I think there is a problem with moses.ini file.
    Could you please send an example .ini file with a sparse feature?

    Thanks
    Arefeh

    On Thu, Aug 18, 2016 at 6:13 PM, Barry Haddow
    <[email protected] <mailto:[email protected]>>
    wrote:

        Hi Arefeh

        The quickest way to see if Moses is using your feature is to
        put a debug message in it to see if it gets called. You can
        also increase the debug of Moses (try -v 2) to see if your
        feature's scores appear in the breakdown.

        To populate the weights file, you will need to run tuning
        (kbmira or pro). If you just decode with an empty weights
        file, all the weights will be set to 0 and Moses will not
        update the file,

        cheers - Barry


        On 18/08/16 14:15, arefeh kazemi wrote:
        Hi Barry

        Thanks.
        I create an empty weights.txt file and write it's address in
        moses.ini. Moses runs normally but weights file remains
        empty. It seems moses doesn't use my feature.

        Regards
        Arefeh

        On Wed, Aug 17, 2016 at 12:57 PM, Barry Haddow
        <[email protected]
        <mailto:[email protected]>> wrote:

            Hi Arefeh

            That seems OK. Tuning (with kbmira or pro) will create a
            weights file for the sparse features, which you can add
            with:

            [weight-file]
            /path/to/sparse/weights

            What goes wrong when you run moses?

            cheers - Barry


            On 17/08/16 07:50, arefeh kazemi wrote:
            Hi
            This is just a kindly reminder that I'm waiting for
            response.

            Thanks
            Arefe
            ---------- Forwarded message ----------
            From: *arefeh kazemi* <[email protected]
            <mailto:[email protected]>>
            Date: Tue, Aug 9, 2016 at 7:28 PM
            Subject: moses.ini file for sparse features
            To: Moses-support <[email protected]
            <mailto:[email protected]>>


            Hi

            I've implemented a sparse feature function in Moses
            Hiero system but I don't know what are the parameters
            in moses.ini file for a sparse feature.
            for the dense version of my feature, I had these lines
            in my ini file:
            [features]
            Aref num-features=4 ...
            [weight]
            Aref= 0.2 0.2 0.2 0.2

            Now, What should I write for my sparse feature in .ini
            file? I have removed the weights for Aref and also
            "num-features" from ini file, but it doesn't work.

            Regards
            Arefeh Kazemi



-- Arefeh Kazemi


            _______________________________________________
            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>
            The University of Edinburgh is a charitable body,
registered in Scotland, with registration number SC005336. -- Arefeh Kazemi
        The University of Edinburgh is a charitable body, registered
in Scotland, with registration number SC005336. -- Arefeh Kazemi
    The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
--
Arefeh Kazemi
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support

Reply via email to