OpenNLP model for model 1.7.3+

["Russ, Daniel (NIH/CIT) [E]" <dr...@mail.nih.gov>]

Hello,

    With the release of OpenNLP 1.7.3, the GISModel serialization will not be backwards compatible with the pre-1.7.3 format.  I am particularly concerned with the models on SourceForge, because I still use them.  They may not be the best models, but they work fairly well and are easily available.  I think we should reformat the old models in the new format, and place the new models on an OpenNLP branded website (e.g Sourceforge, github, apache website).

Daniel

Re: OpenNLP model for model 1.7.3+

[Joern Kottmann <ko...@gmail.com>]

Hello,

I agree with Richard, we can't do such a step in a minor version increase
because we also promise that models work with older minor versions e.g.
model trained with 1.7.4 is supposed to work with 1.7.0.

Users probably have a much higher overhead to retrain their models than to
update to the latest OpenNLP version, for us it is an advantage if our
users can update without retraining (and also frequently do update).

I suggest we only consider breaking model backward compatibility if there
are really good reasons for it.

Jörn

On Sun, Jan 29, 2017 at 10:33 AM, Richard Eckart de Castilho <rec@apache.org
> wrote:

> I think incompatible models would warrant at least an increase in the
> minor version number (1.8.x) if not even in the major version number...
>
> If they can be reformatted from the binary versions (i.e. without
> retraining),
> if would be great if OpenNLP would include a migration tool.
>
> Mind this does not only affect the models on Sourceforge, but also models
> that
> people might have at home or at their institutions but for which they no
> longer
> have the corpus data or there are other reasons why it would be difficult
> for them
> to retrain them...
>
> Cheers,
>
> -- Richard
>
> > On 28.01.2017, at 05:17, Russ, Daniel (NIH/CIT) [E] <dr...@mail.nih.gov>
> wrote:
> >
> > Hello,
> >
> >    With the release of OpenNLP 1.7.3, the GISModel serialization will
> not be backwards compatible with the pre-1.7.3 format.  I am particularly
> concerned with the models on SourceForge, because I still use them.  They
> may not be the best models, but they work fairly well and are easily
> available.  I think we should reformat the old models in the new format,
> and place the new models on an OpenNLP branded website (e.g Sourceforge,
> github, apache website).
> >
> > Daniel
> >
>
>

Re: OpenNLP model for model 1.7.3+

[Richard Eckart de Castilho <re...@apache.org>]

I think incompatible models would warrant at least an increase in the
minor version number (1.8.x) if not even in the major version number...

If they can be reformatted from the binary versions (i.e. without retraining),
if would be great if OpenNLP would include a migration tool. 

Mind this does not only affect the models on Sourceforge, but also models that
people might have at home or at their institutions but for which they no longer
have the corpus data or there are other reasons why it would be difficult for them
to retrain them...

Cheers,

-- Richard

> On 28.01.2017, at 05:17, Russ, Daniel (NIH/CIT) [E] <dr...@mail.nih.gov> wrote:
> 
> Hello,
> 
>    With the release of OpenNLP 1.7.3, the GISModel serialization will not be backwards compatible with the pre-1.7.3 format.  I am particularly concerned with the models on SourceForge, because I still use them.  They may not be the best models, but they work fairly well and are easily available.  I think we should reformat the old models in the new format, and place the new models on an OpenNLP branded website (e.g Sourceforge, github, apache website).
> 
> Daniel
>