Log prior computation

[befelix]

I was going through the GPy code and found some part that seems wrong to me (or that I don't understand) in the computation of the log prior.

The first part of the code computes the log prior probabilities, which makes perfect sense. The second part starting in line 57 supposedly adds some correction terms for the constraints, but I don't understand where they come from. Why would the transformed parameter space matter, when the prior is defined on the untransformed space and param_array returns the untransformed parameters?

Thanks for your help.

[befelix]

Here's roughly what I would have expected:

def log_prior(self):
    """evaluate the prior"""
    if self.priors.size == 0:
        return 0.
    x = self.param_array
    #evaluate the prior log densities
    return sum(p.lnpdf(x[ind]).sum() for p, ind in self.priors.items())

def _log_prior_gradients(self):
    """evaluate the gradients of the priors"""
    if self.priors.size == 0:
        return 0.
    x = self.param_array
    ret = np.zeros(x.size)
    
    #compute derivate of prior density
    for p, ind in self.priors.items():
        np.put(ret, ind, p.lnpdf_grad(x[ind]))
        
    # Transform gradients to unconstrained space
    for c, idx in self.constraints.items():
        if isinstance(c, Transformation):
            ret[idx] = c.gradfactor(x[idx], ret[idx])
    return ret

[mzwiessele]

@jameshensman do you recall why this was necessary?

[mzwiessele]

@jameshensman ping

[JasperKirton]

bumping this, can result in some strange transformations of the log prior