opennlp-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Joern Kottmann <kottm...@gmail.com>
Subject Re: Early stopping NameFinderME
Date Tue, 29 Aug 2017 14:32:44 GMT
Hi Daniel,

do you see any issue if we expose LLThreshold and allow the user to
change it via training parameters?

Jörn

On Sat, Aug 26, 2017 at 1:07 AM, Daniel Russ <danruss00@gmail.com> wrote:
> Jörn,
>
>    Currently, GISTrainer has a private static final variable LLThreshold, which controls
if the change in the log likelihood between two iterations is too small.  We could make this
parameter. I am concerned about using the accuracy to train the model.  If we use accuracy,
the weight space may be flat.
>
>    Saurabh, you use the term “early stopping”.  In deep learning, early stopping
is used to prevent overtraining and improve generalization to unseen data.  I am not sure
early stopping serves the same purpose with GIS training.  Does anyone know if early stopping
improves generalization for a maxent problem?
>
> Daniel
>
>> On Aug 24, 2017, at 4:48 AM, Joern Kottmann <kottmann@gmail.com> wrote:
>>
>> You are the first one who ever asked this question. I think we have this as
>> an option already on the gis trainer but it is not exposed all the way
>> through.
>>
>> Please open a jira and I can look at it next week.
>>
>> Jörn
>>
>> On Aug 21, 2017 5:11 PM, "Saurabh Jain" <saurabh4768jain@gmail.com> wrote:
>>
>>> Hi All
>>>
>>> How can we use early stopping while training/crossvalidating custom data
>>> with NameFinder ? What I want if change in likelihood value or accuracy of
>>> model is less than 0.05 between two steps (differ by 5 i.e compare x+5 step
>>> output with x step) then training should stop. I could not find anything
>>> regarding this in documentation. Can some one please help ?
>>>
>>> --
>>> *Thanks & Regards*
>>>
>>>
>>> *Saurabh Jain *
>>> *AI Developer*
>>>
>>> *Active Intelligence  *
>>>
>>> *"*
>>> *To do a thing yesterday was the best time . Second best time is today .” *
>>>
>

Mime
View raw message