Hi Hieu,
thanks for your answer. I guess that, in case the search is not completely
deterministic, I will be better off by modifying the -osgx option so that
it will output all the info I need. I was trying to avoid that, but it
seems there will be no other option.
Thanks,
Germán
On Wed, 26 Oct 2011, Hieu Hoang wrote:
hi german
I don't think decoding is guaranteed to give the same results in 2 consecutive
runs,
for the same model and input. This email thread was discussing this issue
http://thread.gmane.org/gmane.comp.nlp.moses.user/4327/focus=4341
However, I'm not sure how much different 2 runs can be.
You might wanna talk to Christian Buck who I think is trying to bash the search
graph
output into something more useful
http://article.gmane.org/gmane.comp.nlp.moses.user/4374/
On 24/10/2011 23:05, Germán Sanchis Trilles wrote:
Hello list,
I was trying to output a search graph with both the single model scores
and
the transition value. While the single model scores are produced by the
option -osgx, the transition value is only present when using the option
-osgx. Since I am not doing this myself, I thought that the best choice
would be to output both files by rerunning Moses (since they seem to be
exclusive) and then merge both into a single file. However, I found a
problem when doing so, which is that both files do not have the same
number
of lines (and hence I assume that the original search-graphs did not have
the same amount of edges. So I would have two questions, I guess:
- Why is there information produced by -osg that is not present when
specifying -osgx?
- Why are -osg and -osgx exclusive? Is this really the desired behaviour?
- Finally, if the same moses.ini is specified with the same models, why
are
the search graphs not identical? Are there random initialisations (or
something similar) that account for this behaviour?
the command lines used when observing these results were pretty simple,
i.e. moses -f <moses.ini> -i <file> [-osg|-osgx]
Thanks in advance for your answer,
best regards,
Germán Sanchis-Trilles
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
_______________________________________________
Moses-support mailing list
[email protected]
http://mailman.mit.edu/mailman/listinfo/moses-support
