To be a little more accurate: the anomaly likelihood is currently a post
process and not part of the model. The recommended approach is to save the
AL state separately from the model (e.g. by pickling the class) just as you
mentioned.  That way you do NOT need to pass in 500 data points again.  The
HTM Engine for example, saves the AL state separately from the model and
resurrects that state each time it loads a model.

It would be more convenient if the model included the Anomaly Likelihood
class and they could both be saved with one save statement, but that is not
the current implementation.  It would be a nice improvement to make it part
of the model itself.

--Subutai

On Wed, Feb 24, 2016 at 8:44 AM, Matthew Taylor <[email protected]> wrote:

> Currently, the anomaly likelihood state is not saved when you persist
> a model. That means when you resurrect the model, it will need to see
> 500 data points before you start getting likelihood values again. I've
> created a feature request to address this:
> https://github.com/numenta/nupic/issues/3021.
> ---------
> Matt Taylor
> OS Community Flag-Bearer
> Numenta
>
>
> On Wed, Feb 24, 2016 at 8:17 AM, Alexandra Vasilova
> <[email protected]> wrote:
> > Hello!
> >
> > I started playing around with the nupic opf framework, saving and loading
> > networks. I am using anomaly likelihood class for evaluating the score.
> What
> > should I do if I want to use AnomalyLikelihood class from the point when
> > model was saved? I'm really sorry if this question is covered in some
> > introductory tutorials, but I can't really find this use case in the
> > examples. I suppose that the model saved by the save function from
> model.py
> > should have all the necessary information, right? Or do I need to save
> the
> > anomalyLikelihood object into some pickle and load it every time?
> >
> > Thank you for the answer!
> > Alexandra
>
>

Reply via email to