Karin,

You don't need to swarm if you are looking to get anomaly indications
on scalar input fields. We have already established a decent set of
model params that work for scalar input. All you need to do is update
the "encoders" values to fix the min/max for your data (as well as
entering the correct field names). See an example here:

https://github.com/numenta/nupic/blob/master/examples/opf/clients/hotgym/anomaly/one_gym/model_params/rec_center_hourly_model_params.py

You can use these model params, and update the encoders section
highlighted here:
https://github.com/numenta/nupic/blob/master/examples/opf/clients/hotgym/anomaly/one_gym/model_params/rec_center_hourly_model_params.py#L21-L48

You'll need to change "kw_energy_consumption" to your input field
label, and provide an accurage "minval"/"maxval" for your scalar
encoder configuration.

Note that these model params include "'inferenceType':
'TemporalAnomaly'", which is what makes this an anomaly detection
model instead of a prediction model.

Have you watched this? https://www.youtube.com/watch?v=1fU2Mw_l7ro
---------
Matt Taylor
OS Community Flag-Bearer
Numenta


On Thu, Nov 5, 2015 at 4:03 AM, Karin Valisova <[email protected]> wrote:
> Hello guys,
>
> I've been playing around with the OPF past days and I'm trying to optimize
> the anomaly scores for my data stream. I have read through the code and
> wiki, but I still can't figure out properly how this all works. I am looking
> for Temporal Anomaly and my data stream is just two columns as in the hotgym
> example, one with hourly timestamp and the other containing values.
>
> 1. If I put on swarming, for example by scripts/run_swarm.py with very basic
> .json defining it - without any custom metrics - what are the anomalies I
> get? Those are the raw values - the fraction of activated vs. predicted
> values, right?
>
> 2. Now, if I want to try the different types of anomalies, where should I
> implement the call in my code? Is it enough to write it into the
> model_params.py information and all the following calculations would be
> implemented by the CLAmodel? How does this affect the swarming?
>
> 3. How does metric used affect the anomaly scores? As I understand it, in no
> direct way, right?
>
> Thank you very much for clearing things up a bit for me!
> Hope I haven't missed some basic tutorial with all the explanations, though.
>
> Have a nice day,
> Karin
>
> --
>
> datapine GmbH
> Skalitzer Straße 33
> 10999 Berlin
>
> email: [email protected]

Reply via email to