Closed kislayabhi closed 10 years ago
@karlnapf , yes, this is working fine with the current shogun version locally. But let me correct the problem you have addressed. Also, I will try to add some description.
@karlnapf , I have updated the PR with this new commit. One thing I wanted to know: In gp notebook , while using ML2, you used : inf = ExactInferenceMethod(GaussianKernel(10, taus[i]), feats_train, ZeroMean(), labels_train, GaussianLikelihood()) .
Here the term taus[i] is 32. Will varying this term matter in learning the hyper parameters?
I dont really get the question. Do you mean from where the ML2 optimisation is started? I think its started from current values...
Nice. Cant wait to try this :)
Ok, finally some feedback on this one:
The ML2 button should be documented. Say what happens when this is clicked
The kernel scale should also be a parameter to set. This is m_scale in CExactInferenceMethod
When one presses the ML2 button, the parameters in the boxes should be updated also, not just the plot.
The ML2 button should also work for other kernels, currently it always reverts to the Gaussian one from what I can see from the results
Label of the button should be "Use ML2" (whitespace) button "clear" should be "Clear" capital
The grid that is used to evaluate the regression function should be more fine. Use a few more points
I suggest you add another button next to the train GP one, which is "Show predictive", which creates heatmaps in the same style as the notebook (ask me if unclear) Those are just a 1D Gaussian centered at the mean prediction with variance given by the GP
Sometimes the ML2 fails and just produces a constant prediction, could you investigate why what is?
@karlnapf , Yeah. Will make sure its done once I get some respite from my current work.
@Saurabh7 could you maybe take care of this?
@kislayabhi is busy with his own stuff and this is more part of your project now :)
Some of the above comments were already addressed in the GP classification demo btw
Suggestion: Have a dropdown menu for parameter selection as in GPClassify
Suggestion: Have a FITC sparse GP regression option where one can put the inducing points by hand (right click) see the notebook for how this works
For reference, will work on this.
I couldn't move the data section after the feature one since I am using the loaded data file to demonstrate feature methods and column major format. oops
Just to make sure: Does this smoothly work locally with the latest Shogun version? I cannot test here