Hi Rico,
no I didn't. I chose sorted vectors because of the memory footprint. 
Before it would use more than ten GB for caching for a 10M corpus, now 
around one or two GB. I remember reading somewhere that 
boost::unordered_set has even a worse overhead than std::set concerning 
memory. The speed-up is just a nice side-effect, although I believe the 
main speed problem was the intersection function.

Thanks for the feedback and you fix.

Best,
Marcin

W dniu 22.08.2012 16:29, Rico Sennrich pisze:
> Hi Marcin,
>
> there was a segfault with hierarchical models, but it's working now.
> Have you compared your speedup with that of using unordered_set instead
> of set?
>
> best,
> Rico
>
>> Date: Wed, 22 Aug 2012 12:05:51 +0200
>> From: Marcin Junczys-Dowmunt <[email protected]>
>> Subject: [Moses-support] Phrase table pruning, filter-pt.cpp
>> To: [email protected]
>> Message-ID: <[email protected]>
>> Content-Type: text/plain; charset=UTF-8; format=flowed
>>
>> Hi,
>> I've just pushed a new version of filter-pt.cpp to master. Should be
>> about 2-3 times faster and use way less memory (just replaced sets with
>> sorted vectors). I did not check it with hierarchical models since I do
>> not use them. Could someone take a look if everything is working for those?
>>
>> Cheers,
>> Marcin
> _______________________________________________
> 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

Reply via email to