el-moatasem / moa Goto Github PK

Purpose of code changes on this branch:

Re-use of the same stream in different tasks using the observer pattern.

What steps will reproduce the problem?
1. Create an instance of SingleClassifierDrift from Java code
2. Set its CLI String to "-d DDM -l (bayes.NaiveBayes)"
3. Perform a prequential in a dataset with at least 40k instances

What is the expected output? What do you see instead?
When debug messages are suppressed by a dummy output stream, instead of 
filtering with 'grep' or just showing them,
algorithm runs 3X faster.

What version of the product are you using? On what operating system?
MOA 2012.08 from Maven

Please provide any additional information below.
The messages are about the drift detection levels: N, W and D.
They could be suppressed by an CLI option.

What steps will reproduce the problem?

Create a filtered stream with more than one filter where the filters are order 
dependent. (i.e. one filter depends on another) 

FilterA depends on FilterB

FilteredStream -f FilterA,FilterB

What is the expected output? 
FilterA is applied to the stream first, followed by FilterB.

What do you see instead?
FilterB will be executed first, and fail.

What version of the product are you using? 
moa-20111003

On what operating system?
Windows 7

Purpose of code changes on this branch:

Dynamic arff file format

What steps will reproduce the problem?
1. The problem is always present, but hidden from user.

What is the expected output? What do you see instead?
Budget is expected to be respected.

What version of the product are you using? On what operating system?
MOA 2012.08 Ubuntu 10.10

Please provide any additional information below.
To correct, just put "this.costLabeling++;" on line 196 of 
ActiveClassifier.java.

Budget condition should be tested with future value.
Look at the additional code "+ 1" in the line below.

costNow = (this.costLabeling - this.numInstancesInitOption.getValue() + 1) / 
((double) this.iterationControl - this.numInstancesInitOption.getValue());

Corrected in clone: 
https://code.google.com/r/dpsabc-mostly-active-learning-mods/

Purpose of code changes on this branch:

Use of generics in instances to allow several different examples to feed MOA

Learner<E>

AbstractClassifier extends Learner<Instance>

Reproducing problem:

GaussianEstimator estimator = new GaussianEstimator();
estimator.addObservation(7.0, 2.0);
estimator.addObservation(9.0, 2.0);
System.out.println(estimator.getMean()); // prints 7.5, expected 8.0
System.out.println(estimator.getVariance()); // prints 1.0, expected 1.333

MOA Version: 3a28638d4f55 May 2, 2012

Patch attached.

I am pretty sure the variance calculation in addObservation(GaussianEstimator) 
is also wrong, but I don't know how to fix it.

"For the system to automatically take care of option handling, the  options 
need to be public members of the class, that extend the moa.options.Option 
type."

Line 115 of AbstractOptionHandler.java:
>>> if (fName.endsWith("Option")) {
should be (one of the possible solutions):
>>> if (fType.getName().endsWith("Option")) {

Purpose of code changes on this branch:

Classifiers that can deal with dynamic addition, deletion and modification of 
attributes in the stream

moa.streams.clustering.FileStream

line 191 reads:

numAttsOption = new IntOption("numAtts", 'a',"", removeAttributes.length+1);

should be:

numAttsOption = new IntOption("numAtts", 'a',"", instances.numAttributes() - 
removeAttributes.length);


Leads to incorrect functioning of clustering GUI.

Version: 2012.09-SNAPSHOT

RandomRBFGeneratorEvents generates Instance objects correctly. The problem 
occurs when they are generated via noise. In that case 
weka.core.Instance.toString() throws exception, which is annoying because is 
not possible to use WriteStreamToARFFFile for example.


Here is an example:


import moa.streams.clustering.RandomRBFGeneratorEvents;
import weka.core.Instance;


public class InstanceGenerator {

    private RandomRBFGeneratorEvents randomRBFGeneratorEvents;

    public InstanceGenerator() {
        randomRBFGeneratorEvents = new RandomRBFGeneratorEvents();
        //randomRBFGeneratorEvents.noiseLevelOption.setValue(0);
        randomRBFGeneratorEvents.prepareForUse();
    }

    public Instance next() {
        return randomRBFGeneratorEvents.nextInstance();
    }

    public static void main(String[] args) {
        InstanceGenerator generator = new InstanceGenerator();

        for (int i = 0; i < 10; i++) {
            Instance instance = generator.next();
            System.out.println(instance.toString());
        }
    }

}

What steps will reproduce the problem?
1. Connect to "Iris" collection in Mongo using Java API. Get DBCursor object 
that points to whole learning set. 

2.Initialize dataSet with "schema" from first Mongo document from the 
collection. Treat every Attribute as nominal (use DBCursor.distinct()) or 
string type.

3. Declare NaiveBayes and initialize it by prepareForUse().

4. Convert BSONs to Instance using BSON.toMap() and SparseInstance(double 
weight, double[] atrValues) constructor with given attributes values that match 
their value index in dataSet. Use setDataset() to give reference for Attributes.

5. Train NaiveBayes by NaiveBayes.trainOnInstanceImp() or 
NaiveBayes.trainOnInstance().

6. Print out NaiveBayes.getVotesForInstance(inst).

What is the expected output? What do you see instead?
There should be probabilities for 3 Iris classes that appears non-Random and 
greater than numeric 0. 

Instead I get values that differ although are 'very small' (i.e. E-15) for 
every class value, Basic(...)Evalutator gives values of 0 for correctly 
classified and the same for incorrectly classified.  

What version of the product are you using? On what operating system?
Linux Ubuntu 12.04.5 / MOA 2014.04a