hiya
Hieu Hoang
http://www.hoang.co.uk/hieu
On 13 October 2016 at 15:08, Vito Mandorino <
vito.mandor...@linguacustodia.com> wrote:
> We haven't checked the probingpt + minlexr speedup yet, however we have
> found some further differences in the output with respect to the standard
> Moses
We haven't checked the probingpt + minlexr speedup yet, however we have
found some further differences in the output with respect to the standard
Moses decoder.
It happens sometimes that the order of replacement of placeholders with
actual numbers is not the good one. For instance :
moses2
Yes I modified the line in the moses.ini . My comparison was with respect
to probingPT + minlexr reordering model (rather than .gz reordering model)
2016-10-07 16:25 GMT+02:00 Hieu Hoang :
> weird. it should be a massive speedup (~500%). You have to change the
> moses.ini
yep, it should give you a big speedup compared to probingpt + minlexr model
Hieu Hoang
http://www.hoang.co.uk/hieu
On 7 October 2016 at 16:21, Vito Mandorino <
vito.mandor...@linguacustodia.com> wrote:
> Yes I modified the line in the moses.ini . My comparison was with respect
> to probingPT +
weird. it should be a massive speedup (~500%). You have to change the
moses.ini file slightly
[feature]
LexicalReordering … path=reordering-table.msd-bidirectional-fe.0.5.0-0.gz
to
[feature]
LexicalReordering … property-index=0
Hieu Hoang
http://www.hoang.co.uk/hieu
On 7 October 2016
Yes, that worked for me as well, thank you. There is a little improvement
in speed but not that much actually (about 5% faster using 30 threads).
2016-10-04 11:44 GMT+02:00 Hieu Hoang :
> yes - the script expects the files to be gzipped.
> It runs ok for me. I executed this:
yes - the script expects the files to be gzipped.
It runs ok for me. I executed this:
MOSES_DIR=~/workspace/github/mosesdecoder.perf
$MOSES_DIR/scripts/generic/binarize4moses2.perl
--phrase-table=phrase-table.gz
--lex-ro=reordering-table.wbe-msd-bidirectional-fe.gz
if guessed you ran binarize4moses2.perl so please give me the phrase-table
and reordering model, and the exact command you ran
Hieu Hoang
http://www.hoang.co.uk/hieu
On 3 October 2016 at 15:20, Vito Mandorino <
vito.mandor...@linguacustodia.com> wrote:
> I have managed to replicate the issue on
I have managed to replicate the issue on a smaller corpus. Do you need the
training corpus, the tables (phrase- and reordering-), or all of them?
Vito
2016-09-30 13:30 GMT+02:00 Hieu Hoang :
> wow, that looks like a seriously problem.
>
> I've not seen this before. If you
I tried the following command:
perl /home/Moses/mosesdecoder/scripts/generic/binarize4moses2.perl
--phrase-table=/home/vito/phrase-table.sorted
--lex-ro=/home/vito/reordering-table.sorted
--output-dir=/home/vito/integrated_phrase-reordering/ --num-lex-scores=6
but it gets stuck with the
use the script:
scripts/generic/binarize4moses2.perl
It takes as input the (text) phrase-table and the (text) lexro model. It
will give you the probing pt which contains the info for both.
To use this script, Moses MUST be compiled with the flag --with-cmph. Also,
the program in
you will get another big speedup fromm integrating the lexro into the pt
Hieu Hoang
http://www.hoang.co.uk/hieu
On 29 September 2016 at 15:03, Vito Mandorino <
vito.mandor...@linguacustodia.com> wrote:
> Yes the model includes a lexicalised reordering model but is not
> integrated into the
Ok thank you, I'll check that. Do you know how to perform the integration?
Juxtaposing the 4 phrase-table scores and the 6 reordering scores before
calling CreateProbingPT2 would be enough?
(I have used the CreateProbingPT2 binary and not CreateProbingPT so far)
Vito
2016-09-29 16:07 GMT+02:00
Yes the model includes a lexicalised reordering model but is not integrated
into the probingPT. The size of the LM is 1.8G.
2016-09-29 15:59 GMT+02:00 Hieu Hoang :
> ps. how big is your LM?
>
> Hieu Hoang
> http://www.hoang.co.uk/hieu
>
> On 29 September 2016 at 14:58, Hieu
great, thanks. Do you use the lexicalised reordering model, and is it
integrated into the phrase-table in Moses2?
There is latency in communicating with the server. As Moses2 is much faster
now, the client can't feed it fast enough. You should see that moses2
command line will max out the CPU,
ps. how big is your LM?
Hieu Hoang
http://www.hoang.co.uk/hieu
On 29 September 2016 at 14:58, Hieu Hoang wrote:
> great, thanks. Do you use the lexicalised reordering model, and is it
> integrated into the phrase-table in Moses2?
>
> There is latency in communicating with
Yes, here are some data:
Average source sentence length: 29 tokens
Phrase-table size, probingPT: 11G
Phrase-table size, compact phrase-table: 2.1G
Translation time Moses2 with 32 threads: 1m36.511s
Translation time Moses with 32 threads: 6m14.248s
Translation time Moses2 with 32 threads in
cool. do you have any indications of speed, especially when using
multiple threads? model sizes and average input sentence length are also
relevant.
___
Moses-support mailing list
Moses-support@mit.edu
Now it works! Thanks. On 6000 test sentences the Moses2 output is now
actually 100% identical to the standard Moses output.
Vito
2016-09-28 16:12 GMT+02:00 Hieu Hoang :
> hi Vito,
>
> please git pull and try decoding again. I've just pushed a fix
>
hi Vito,
please git pull and try decoding again. I've just pushed a fix
https://github.com/hieuhoang/mosesdecoder/commit/0005e98b2674906162ce7945c5edd6a42c9ca418
Basically, I've changed changed the behavious of the pugi call so that it
doesn't unescape the words
Hieu Hoang
ah ok. do you have a moses.ini and example input sentence to go with that.
pugixml.cpp is used to parse the input sentence for XML markups for
placeholders, forced-translation etc. You shouldn't change the code for
pugixml 'cos it's an imported library that we don't control and we may
reimport in
We are able to replicate the issue with the probingPT version of this
phrase-table:
||| ||| 1 1 1 1 ||| 0-0 ||| 1 1 1 ||| |||
||| ||| 1 1 1 1 ||| 0-0 ||| 1 1 1 ||| |||
||| ||| 1 1 1 1 ||| 0-0 ||| 1 1 1 ||| |||
||| ||| 1 1 1 1 ||| 0-0 ||| 1 1 1 ||| |||
||| ||| 1 1 1 1 ||| 0-0 ||| 1 1 1
Can you make your model files available for download?
Moses and Moses2 aren't guaranteed to give exactly the same answer.
However, they should be the same quality overall
Hieu Hoang
http://www.hoang.co.uk/hieu
On 28 September 2016 at 09:53, Vito Mandorino <
vito.mandor...@linguacustodia.com>
23 matches
Mail list logo