David,
Thanks for the quick response. Now that I look at the code I should’ve
guessed that behavior. However, won’t that create a lot of overhead with
projects that have many input fields but are interested in a single classifier?
Loving the observer pattern makes output and debugging very easy.
Also, I’m using the Publisher/ObservableSensor feature for the
pre-editing phase — pre-sensor processing, data munging etc... Having all of
this in one executable makes things much easier to maintain and execute. In the
past I had a separate process to work the data into manageable form prior to
sending it to the HTM.
Thanks
-Phil
> On Dec 22, 2015, at 9:36 PM, cogmission (David Ray)
> <[email protected]> wrote:
>
> Hi Phil,
>
> Welcome! and thanks for using HTM.java! It's great to hear from people who
> are using it.
>
> When setting up the network (NAPI) , and when using the line:
>
> Network network = Network.create("test network", p)
> .add(Network.createRegion("r1")
> .add(Network.createLayer("1", p)
> ----> .alterParameter(KEY.AUTO_CLASSIFY, Boolean.TRUE)
> .add(Anomaly.create())
> .add(new TemporalMemory())
> .add(new SpatialPooler())
> .add(Sensor.create(FileSensor::create,
> SensorParams.create(
> Keys::path, "",
> ResourceLocator.path("rec-center-hourly.csv"))))));
> ==============
>
> ... alterParameter(KEY.AUTO_CLASSIFY, Boolean.TRUE)
>
> Classifiers are automatically set up for you. At this time, the NAPI sets up
> classifiers for all fields (each with its own classifier) - without the
> ability to limit the classifier creation to one or some number of classifiers
> less than the number of fields.
>
> So no, you do not have to worry about which column your intended field
> resides in terms of sequence order.
>
> Classifiers and their results may be retrieved using the following syntax:
>
> (From your subscriber...)
>
> public void onNext(Inference i) {
> Map<String, NamedTuple> m = i.getClassifierInput();
> System.out.println(m);
> }
>
> ...will print out the Encoder inputs used to associate Classifier inputs,
> such as:
>
> key=encoding, value=[I@32fa1633, hash=3
> key=bucketIdx, value=20, hash=3 <-- bucket here refers to encoder
> bucket
> key=name, value=timestamp, hash=3
>
> =========
>
> public void onNext(Inference i) {
> NamedTuple nt = i.getClassifiers();
> System.out.println(nt);
> }
>
> ...will print out the NamedTuple containing the field name key to classifier
> values, such as:
>
> keys: Bucket: 0
> Bucket: 1
> key=consumption,
> value=org.numenta.nupic.algorithms.CLAClassifier@7d907bac, hash=1
> Bucket: 2
> key=timestamp,
> value=org.numenta.nupic.algorithms.CLAClassifier@7791a895, hash=2
> Bucket: 3
>
> ... The above Buckets refer to the NamedTuple hash buckets not the encoder
> (just pointing that out).
>
> =========
>
> public void onNext(Inference i) {
> System.out.println("keys c: " +
> i.getClassification("consumption").getMostProbableValue(1));
> System.out.println("keys t: " +
> i.getClassification("timestamp").getMostProbableValue(1));
> }
>
> ... will print out the most probable values as computed by the two
> classifiers (how many classifiers you have will correspond to the number of
> fields).
>
> example printout:
>
> keys c: 21.2
> keys t: 2010-07-02T00:00:00.000-05:00
> keys c: 16.4
> keys t: 2010-07-02T00:00:00.000-05:00
> keys c: 16.4
> keys t: 2010-07-02T01:00:00.000-05:00
> keys c: 4.7
> keys t: 2010-07-02T02:00:00.000-05:00
> keys c: null
> keys t: null
> keys c: 4.67
> keys t: null
> keys c: 23.5
> keys t: 2010-07-02T05:00:00.000-05:00
> keys c: 23.5
> keys t: 2010-07-02T05:00:00.000-05:00
> keys c: 45.4
> keys t: 2010-07-02T07:00:00.000-05:00
> keys c: null
> keys t: null
>
>
> nulls are produced in some cases because I only ran this for 9 iterations so
> the network didn't have time to "warm up".
>
> ============
>
> At this time, all fields must have corresponding Field Encoding Maps setup
> and inserted into your Parameters. For example:
>
> Map<String, Map<String, Object>> fieldEncodings =
> NetworkTestHarness.getNetworkDemoFieldEncodingMap();
>
> Parameters p = Parameters.getAllDefaultParameters();
> p.setParameterByKey(KEY.FIELD_ENCODING_MAP, fieldEncodings);
>
> The NetworkTestHarness is meant to be used as a convenience guide for
> parameters setup. The defaults it uses will not work for all cases, but it
> shows the technique by which you would enter parameters for each one of your
> columns (an unfortunate necessity at this point in time).
>
> I hope this helps!
>
> Cheers,
> David
>
>
> On Tue, Dec 22, 2015 at 7:11 PM, Phil iacarella <[email protected]
> <mailto:[email protected]>> wrote:
> Hey Nupic,
>
> I’m in the process of using NuPIC to predict the next fueling location of
> drivers. I’ve setup a single 2/3 layer with SP and Classifier. I’m
> implementing in HTM.java. The data being fed into the layer consists of 3
> fields (there could be more fields in the future depending on the relevance):
> datetime, driverID, Location.
>
> My question is how do I specify the field to be used in the Classifier. It
> looks as if the Classifier assumes the 2nd field to be the one to classify
> by. Do I have to organize my inputs such that the classified field is in the
> 2nd position?
>
> Thanks
> Phil
>
>
>
> --
> With kind regards,
>
> David Ray
> Java Solutions Architect
>
> Cortical.io <http://cortical.io/>
> Sponsor of: HTM.java <https://github.com/numenta/htm.java>
>
> [email protected] <mailto:[email protected]>
> http://cortical.io <http://cortical.io/>
