Multi Step Predictions in Java

wip_user · July 3, 2017, 6:43pm

Hi
Pls let me know how to configure java code to get future predictions . suppose I want to know next 6 predictions . I found lot of examples in Python not in Java

wip_user · July 4, 2017, 9:42pm

It seems current HTM Java version is not giving option to developer to pass number of steps ,
the below piece of code in Layer.java limiting multi step prediction in java , I made some change to in order to work.

ca[i] = new SDRClassifier();

  /**
     * Creates the {@link NamedTuple} of names to encoders used in the
     * observable sequence.
     * 
     * @param encoder
     * @return
     */
    @SuppressWarnings("unchecked")
    NamedTuple makeClassifiers(MultiEncoder encoder) {
        Map<String, Class<? extends Classifier>> inferredFields = (Map<String, Class<? extends Classifier>>) params.get(KEY.INFERRED_FIELDS);
        if(inferredFields == null || inferredFields.entrySet().size() == 0) {
            throw new IllegalStateException(
                    "KEY.AUTO_CLASSIFY has been set to \"true\", but KEY.INFERRED_FIELDS is null or\n\t" +
                    "empty. Must specify desired Classifier for at least one input field in\n\t" +
                    "KEY.INFERRED_FIELDS or set KEY.AUTO_CLASSIFY to \"false\" (which is its default\n\t" +
                    "value in Parameters)."
            );
        }
        String[] names = new String[encoder.getEncoders(encoder).size()];
        Classifier[] ca = new Classifier[names.length];
        int i = 0;
        for(EncoderTuple et : encoder.getEncoders(encoder)) {
            names[i] = et.getName();
            Class fieldClassifier = inferredFields.get(et.getName());
            if(fieldClassifier == null) {
                LOGGER.info("Not classifying \"" + et.getName() + "\" input field");
            }
            else if(CLAClassifier.class.isAssignableFrom(fieldClassifier)) {
                LOGGER.info("Classifying \"" + et.getName() + "\" input field with CLAClassifier");
                ca[i] = new CLAClassifier();
            }
            else if(SDRClassifier.class.isAssignableFrom(fieldClassifier)) {
                LOGGER.info("Classifying \"" + et.getName() + "\" input field with SDRClassifier");

                ca[i] = new SDRClassifier();
            }
            else {
                throw new IllegalStateException(
                        "Invalid Classifier class token, \"" + fieldClassifier + "\",\n\t" +
                        "specified for, \"" + et.getName() + "\", input field.\n\t" +
                        "Valid class tokens are CLAClassifier.class and SDRClassifier.class"
                );
            }
            i++;
        }
        return new NamedTuple(names, (Object[])ca);
    }

I had to make some change in the

rhyolight · July 4, 2017, 10:09pm

@wip_user Today is a holiday (actually it has been a 4-day weekend for many) in the US so @cogmission probably won’t see this for a bit.

wip_user · July 5, 2017, 3:21am

No Issues , I can wait long I dont want to disturb anyone holidays including your holidays

cogmission · July 5, 2017, 10:57am

Hi @wip_user,

(Thanks @rhyolight, I was indeed AFK for a couple of days)

It’s true that the current version of HTM.Java’s Network API doesn’t offer the setting of more than one step ahead with predictions. At the time, it wasn’t seen as critical feature to support - but if you’d like, you can create an issue in the project and I can take a look at it and add the feature? Or you can fill out a Contributor’s License and submit a Pull Request with your own work? It can be very rewarding to contribute to open source, if you haven’t done it before… It’s up to you - I’m happy to work on it if you’d rather that?

Also, if anyone else would like to work on that feature, it should be fairly straightforward and could possibly lead to Fame and Fortune!

wip_user · July 5, 2017, 10:07pm

Thanks David , currently I have done in crude way , I will find out right way of doing .

Thanks & Regards
MH

Matheus_Araujo · September 20, 2018, 6:18pm

Please, tell how could I contrbute. Could you point me the directions? I’d be glad to contribute.

Matheus_Araujo · October 16, 2018, 12:33pm

Cogmission, I think the main change I have to do is bellow:

Replace this line

ca[i] = new SDRClassifier();

With this one:

ca[i] = new SDRClassifier(new TIntArrayList(steps), 0.001, 0.3, 0);

Now, my doubt is: where do I find the inferred field steps configuration in order to fill the step array?

Matheus_Araujo · October 16, 2018, 12:45pm

I think I will have to create a new parameter in Parameter.java like this:

INFERENCE_STEPS(“inferenceSteps”, int[].class),

Matheus_Araujo · October 16, 2018, 1:19pm

It is working now!

MyNetworkHarness.java

p.set(KEY.INFERRED_FIELDS, getInferredFieldsMap(inferenceField, SDRClassifier.class));
p.set(KEY.INFERRED_STEPS, new int[]{1, 2, 3, 10, 100});

`

defaultEncoderParams.put(KEY.INFERRED_STEPS, new int{1});

`

The new method on Layer.java

@SuppressWarnings("unchecked")
    NamedTuple makeClassifiers(MultiEncoder encoder) {
        Map<String, Class<? extends Classifier>> inferredFields = (Map<String, Class<? extends Classifier>>) params.get(KEY.INFERRED_FIELDS);
        int[] steps  = (int[]) params.get(KEY.INFERRED_STEPS);
        if(inferredFields == null || inferredFields.entrySet().size() == 0) {
            throw new IllegalStateException(
                    "KEY.AUTO_CLASSIFY has been set to \"true\", but KEY.INFERRED_FIELDS is null or\n\t" +
                    "empty. Must specify desired Classifier for at least one input field in\n\t" +
                    "KEY.INFERRED_FIELDS or set KEY.AUTO_CLASSIFY to \"false\" (which is its default\n\t" +
                    "value in Parameters)."
            );
        }
        
        String[] names = new String[encoder.getEncoders(encoder).size()];
        Classifier[] ca = new Classifier[names.length];
        int i = 0;
        for(EncoderTuple et : encoder.getEncoders(encoder)) {
            names[i] = et.getName();
            Class fieldClassifier = inferredFields.get(et.getName());
            if(fieldClassifier == null) {
                LOGGER.info("Not classifying \"" + et.getName() + "\" input field");
            }
            else if(CLAClassifier.class.isAssignableFrom(fieldClassifier)) {
                LOGGER.info("Classifying \"" + et.getName() + "\" input field with CLAClassifier");            
                ca[i] = new CLAClassifier(new TIntArrayList(steps), 0.001, 0.3, 0);
            }
            else if(SDRClassifier.class.isAssignableFrom(fieldClassifier)) {
                LOGGER.info("Classifying \"" + et.getName() + "\" input field with SDRClassifier");
                ca[i] = new  SDRClassifier(new TIntArrayList(steps), 0.001, 0.3, 0);
            }
            else {
                throw new IllegalStateException(
                        "Invalid Classifier class token, \"" + fieldClassifier + "\",\n\t" +
                        "specified for, \"" + et.getName() + "\", input field.\n\t" +
                        "Valid class tokens are CLAClassifier.class and SDRClassifier.class"
                );
            }
            i++;
        }
        return new NamedTuple(names, (Object[])ca);
    }

Matheus_Araujo · October 16, 2018, 1:20pm

Now, how do I submit this code to the github project?

cogmission · October 16, 2018, 2:57pm

@Matheus_Araujo,

Hi,

Congrats for getting your code working! That’s very exciting!

In order to submit code, there are a few steps… I will start from the assumption however that you are already familiar with Git, because the subject of git management is a lengthy conversation.

You will need to create an issue in the project.
You will have to fill out the Numenta Contributor’s License which will then be approved by @rhyolight (Matt Taylor).
You will need to fork the project, and create a branch off of “Main” to submit your code.
You then commit your code on that branch and push it up to your repository.
You then create a PR (Pull Request) and in the topic you put the words “Fixes #”. (see other PR’s under closed to see how it looks)

It’s a shorter process than what it looks like here. Only made longer because this is your first time…

Good luck! Let me know if you have any other questions?

Cheers,
David

Matheus_Araujo · October 16, 2018, 3:20pm

cogmission, I had created created an issue long ago and I have signed th contributor license too. Now I created a new pull request but haven’t done the FIX# thing…

And I see now that I hae sent some modifications I made to some files in order to have it running under my corporate proxy through cntlm proxy tunneling.

My pull request: https://github.com/numenta/htm.java/pull/544

My fork: https://github.com/matheus-fatguys/htm.java

My Issues:

Sorry, but this is my first free software contribution and I’m not used to it.

Matheus_Araujo · October 16, 2018, 3:25pm

I’ve updated the PR comment and now it is as bellow:

Fixes #543 #542
Enabling the Network API to create multi step inferences and fixing a bug in multiencoder that prevented one from using coordinate encoder

Topic		Replies	Views
Setting step in a classifier HTM.Java	7	850	January 17, 2019
New HTM.Java Release! (v0.6.12-alpha) HTM.Java	4	1729	April 13, 2017
Is HTM multi-step prediction currently possible? Numenta Theory	5	1580	May 22, 2018
Anomalies vs predictions HTM.Java	1	854	May 17, 2017
HTM.java network setup for making predictions HTM.Java	8	1502	February 9, 2017

Multi Step Predictions in Java

Related topics