hi tom
good questions
On 02/06/2012 16:23, Tom Hoar wrote:
> Hi Hieu,
>
> Hope all is well with you where ever you hang hour had now-days.
> Before updating my Github to these changes, I reviewed the previous
> train-model.perl. It generated some questions and a couple
> suggestions, but sending them direct vice moses-support.
>
> 1) in the previous version, the value of $BINDIR was set during make
> install. Users had the -bin-dir argument to override the default. How
> is your -external-bin-dir update different?
-external-bin-dir is $BINDIR from previous versions. However, i don't
set a default value for it. it must be set from the command line.
>
> 2) The mgizapp setup that sets $SNT2COOC seems to be in error when it
> says "# Since "snt2cooc.out" and "snt2cooc" work the same". The last I
> checked, they do not have the same command-line interface. I'm still
> working with MGIZA++ SVN rev 16 because anything newer is broken when
> using the repository version of boost libraries. I've reported the
> problems, but seems Qin is pretty busy. Everything from rev 17 on
> comes with a disclaimer it's in transition to work on MS Windows and
> Linux might be broken. So, I don't know if the newer updates changed
> the snt2cooc command line.
new mgiza works for me. snt2coocis annoyingly slightly different between
giza++ & mgiza. That's taken care of in
train-model.perl line 1000
If there are problems with them, report it to the original developers,
or to the moses mailing list if you have to.
> 3) train-model.perl default setting for $_MGIZA_CPUS is hard-coded to
> 4. How about using something like this to set it based on the host's
> resources. I don't know if Sys:CPU is a standard Perl resource.
>
> use Sys::CPU;
> $number_of_cpus = Sys::CPU::cpu_count();
its not installed by default on my mac. Not sure about unixes. May be
more trouble than its worth. Unless you wanna commit it and deal with
the tantrums
>
> 4) there was a moses-support thread a while back that said MGIZA++'s
> merge_alignment.py is broken for CPU counts over 10. I checked into it
> and sure enough, MGIZA++ uses a strange way to increment its temp file
> numbering. It increments based on the ASCII decimal value starting at
> 48 (or 49 don't remember), so after '0' (or '1') thru '9', filenames
> change to ':', ';', '<', etc. I was going to update the python script
> to count the same, but haven't had time find the exact string
> concatenation. Maybe this next week or two.
look forward to the fix.
>
> Thanks,
> Tom
>
>
> On Thu, 31 May 2012 20:42:56 +0100, Hieu Hoang
> <fishandfrol...@gmail.com> wrote:
>> Hi all
>>
>> if you're checking out the latest github code, there are some changes
>> you should be aware of:
>> 1. There is a new argument to train-model.perl
>> -external-bin-dir [path]
>> This points to the directory where Giza++/mgiza lives. Setting
>> this is MANDATORY if you're using train-model.perl to do the word
>> alignment. It used to be hardcoded in the perl code itself.
>> 2. All the training programs have been moved into the directory
>> [MOSES-ROOT]/bin
>> They should be run from there, not from wherever the source
>> code is.
>> 3. To roll out, simply copy the 2 directories
>> [MOSES-ROOT]/bin
>> [MOSES-ROOT]/scripts
>> to wherever you want, eg.
>> /home/hieu/moses/bin
>> /home/hieu/moses/scripts
>> 4. If you don't want to move it anywhere, you can run it from where
>> you downloaded.
>> 5. The EMS and example files have been updated.
>>
>> Hope this is ok for everyone. It may break some people's setup. If
>> possible, please change your setup. It's gonna help us all in the long
>> run. If not, flame me & i'll see what I can do
>>
>> HH
>>
>>
>> _______________________________________________
>> Moses-support mailing list
>> Moses-support@mit.edu
>> http://mailman.mit.edu/mailman/listinfo/moses-support
>
>
_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support
