I have an epsilon-SVR model in which I trained, with a specific epsilon of say 0.001. I get the following output,
[LibSVM]*.*
optimization finished, #iter = 69
obj = -0.032164, rho = 0.006072
nSV = 11, nBSV = 4
From the output I know that 4 of my support vectors are bounded i.e. lay on the margin.
Now the answer is simple of course, just
numpy.argwhere(svr_model.predict(svr_model.support_vectors) == svr_model.epsilon)
Or possibly
numpy.argwhere(trainY(svr_model.support_) == svr_model.epsilon)
Yet either way I return none. I understand that it is to do with the decimal points that come after the predictions but what is the tolerance for this?
Am I even doing this right?