ARD and RVR: variance of predicitve distribution, take inverse?

[sophiegif]

Hello,
it seems that in the predict_dist functions, the variance output is calculated using this formula that seems true to me:
var=sigma^2+(K^T Sigma K)

In the code, sigma^2 (identified as alpha_) is directly the precision beta (in the fit method, self.alpha_ = beta). But the precision is usually the inverse of the variance (and seems to be calculated as the inverse of the variance, even in your initialization).

I think the code is good, just the predict_dist should be :
var_hat = 1./self.alpha_ ( and not var_hat = self._alpha)
var_hat += ...

If my thoughts are false, please do correct me!
Thank you again, I know looking back to your code must be time consuming...

[AmazaspShumik]

I will need to double check, later today, but it seems you are right. (I assume you are speaking about code in fast_rvm.py, right ? ).
Can you make pull request?

[sophiegif]

Yes, taking about fast_rvm.py
I just did pull request.

Thanks!

[AmazaspShumik]

Cool )

[AmazaspShumik]

Thanks for help, I accepted pull request