Implement log-likelihood for test set

[AmazaspShumik]

Since optimization is non convex we can end up in local maximum. So to find best fit we may need to rerun algorithm several times therefore we need a method to compare goodness of fit of two models.