Hi,
I think the answer to the performance problem is to only track the
total score during decoding. Then, if feature values are requested,
reconstruct them from the derivation of each n-best list entry. Perhaps
somebody is looking for a project before or during MT Marathon.
Kenneth
On 05/31/2012 04:49 PM, Tetsuo Kiso wrote:
> Hi Barry,
>
> I see.
> It seems that there're couple of things to try out such as feature
> selection or the size of tuning dataset before merging.
>
> Tetsuo
>
> On Jun 1, 2012, at 4:19 AM, Barry Haddow wrote:
>
>> Hi Tetsuo
>>
>> Sorry about the delay!
>>
>> Yes, I'd like to merge the miramerge back into to trunk. At the moment
>> though,
>> I have concerns about the performance of the sparse features (time and
>> space),
>> so I'd like to address these before the merge. I'm going to have a look at it
>> over the summer, and probably solicit help at the mt marathon,
>>
>> cheers - Barry
>>
>> On Saturday 26 May 2012 15:03:48 Tetsuo Kiso wrote:
>>> Hi Barry,
>>>
>>>> I've updated the sparse features branch of Moses (miramerge), merging all
>>>> the recent changes from trunk. It passes all the unit tests, and all the
>>>> regression tests (apart from small number of rounding errors) but I
>>>> have't run any other tests on this branch.
>>>
>>> I'm just wondering is there any plan to merge the branch "miramerge"
>>> into "master" in the future?
>>>
>>>
>>> Tetsuo
>>>
>
>
> _______________________________________________
> 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
