gravitylab / edda Goto Github PK

joint_GMM and joint_gaussian are updated and some interfaces are changed.
The python wrapper may have minor problem now.
This has to to be verified.
I do not have the environment to test this now.
Someone can help this or I will do this later.

Shall we add this function? Or use eddaComputeJointGMM with one component whenever needed?

log_likelihoods(smpPtr, 1) = 0;

should it be log_likelihoods(smpPtr, 0) ?
similarly some other log_likelihoods(smpPtr, 1) statements

When checked with the GMM estimation of OpenCV lib the corresponding edda version gives very different results (possibly wrong). The weights of all the Gaussian components seems to have equal values.

We can discuss this issue later.

EM sometimes never converges. So a maximum iterations have to be set.
In the current GMM training version, the maximum iterations is set to 100.
But we can discuss other choice.
For example, (1) let users input the maximum number of iterations (2) determining the maximum number of iterations by analysing the input samples (based on some algorithms)

Check the wrapper JointHistogram::marginalization() in pyEdda.py and py_joint_histogram.h

There might be potential name conflict in the c-like functions we wrapped for different classes, e.g. getJointMean() for JointHistogram and for JointGaussian.