07:12 < marcus_zoq> udit_s: Just for clarification, you transform the target to a 2D-vector if we use two labels and the output is a 3D-vector if we use 3 labels, right?

08:03 < udit_s> marcus_zoq: 3D-vector as in a cube ? for n class labels, I just have a weight matrix with n rows.

08:04 < udit_s> marcus_zoq: also, I think I've figured out what the problem was and why yesterday's solution wasn't converging.

08:04 < udit_s> - no bias vector.

08:04 < udit_s> I'm taking care of that right now.

08:19 < marcus_zoq> udit_s: If you compute 'weightVectors * data.col(j);' I was expecting a value, so I though you transform the target value to a binary vector.

08:21 < marcus_zoq> udit_s: Something like: y = x(i,1)*w(2,1)+x(i,2)*w(3,1) -> single value

08:26 < udit_s> No, I'm getting the max value from tempLabelMat. as the single value.

08:27 < udit_s> marcus_zoq: Oh, and I think I fixed it. Just need to run some more tests to be sure, but it works for AND, OR gate implementations and the random linearly separable test case

08:27 < udit_s> Let me just upload the code.

08:32 < udit_s> udit_s: So basically, I've added a col in weightVector for the bias vector. And inserted a row of ones in trainData and testData for bias variables. It properly converges now.

08:33 < udit_s> marcus_zoq: ^ oops.

08:40 < marcus_zoq> udit_s: Okay sounds good?

08:45 < udit_s> marcus_zoq: So what next ? I'll write a few more tests, edit the documentation; are we assuming linearly separable data as the input ?

08:53 < marcus_zoq> udit_s: Yeah we are assumig linearly separable data. Test great idea, and a main would be great.

08:55 < marcus_zoq> udit_s: Can you also add a test with just two inputs?

15:58 < Anand> Marcus : I didn't really get you. What is the m_rates vector? How do I implement the get_probs() method?

16:52 < marcus_zoq> Anand: Hello, here is the shogun nbc header file and the cpp file:

16:52 < marcus_zoq> https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/GaussianNaiveBayes.h

16:52 < marcus_zoq> https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/multiclass/GaussianNaiveBayes.cpp

16:53 < marcus_zoq> What we need is the m_rates vector. The problem is, we can't directly access this vector. A protected member is accessible in the class that defines them and in classes that inherit from that class. So, we create a new class that inherit form the shogun CGaussianNaiveBayes class.

16:54 < Anand> Yeah right, I got that. But, what exactly is m_rates?

16:55 < Anand> log of m_prob?

16:55 < marcus_zoq> Anand: A shogun vector object that contains the probabilities.

16:56 < marcus_zoq> Anand: m_rates.vector[i] = CMath::log(m_label_prob.vector[i]);

16:56 < Anand> yes, exactly

16:56 < Anand> why the log?

16:58 < Anand> And later m_rates.vector[i]+= CMath::log(0.39894228/CMath::sqrt(m_variances(k, i))) - 0.5*CMath::sq(feature_vector.vector[k]-m_means(k, i))/(m_variances(k, i));

17:13 < marcus_zoq> Anand: 'why the log' -> The class with the highest log probability value is the most probable class.

17:13 < marcus_zoq> Anand: 'm_rates.vector[i]+= CMath::log( ...' -> This should be the equation for the normal distribution.

17:14 < marcus_zoq> Anand: Let me search for a formula.

17:14 < marcus_zoq> Anand: http://en.wikipedia.org/wiki/Naive_Bayes_classifier#Gaussian_naive_Bayes

17:15 < marcus_zoq> Anand: We can compare the results with mlpack or matlab.

17:16 < Anand> Ok. Now, I get this

17:17 < Anand> So, we want to call the apply_one(...) function and set the values in m_rates for each instance. Right?

17:18 < Anand> This gives probabilities for a single instance at an index in the file. Correct?

17:19 < marcus_zoq> Anand: You can use apply_multiclass(...) this calls apply_one for the dataset and afterwards you can look into the m_rates vector.

17:20 < Anand> But, m_rates is for a single instance right? It is a one dimensional vector

17:25 < Anand> So, we will write the m_rates vector to file after each call to apply

17:25 < Anand> -_one

17:25 < marcus_zoq> Anand: Yeah I think you are right!

17:27 < marcus_zoq> Anand: Your are right!

17:27 < Anand> Ok. I will add the get_probs(..) method

17:27 < Anand> After that, how do we use it in python nbc.py?

17:29 < Anand> we will need an executable and then we will run it just as we did in weka. This will give us the required files

17:29 < marcus_zoq> Anand: We need to write a small c++ main/executalbe that uses the new class.

17:29 < marcus_zoq> Anand: Right :)

17:29 < Anand> Ok. I think I can do it now.

17:30 < marcus_zoq> Anand: I think, you can use the function from the last mail as a basis.

17:31 < Anand> Yeah, sure

17:32 < marcus_zoq> Anand: But you need to install shogun or use the build server to test the code

17:33 < Anand> the python module?

17:33 < marcus_zoq> Anand: The c++ library.

17:34 < Anand> Ok. I will install it