I’m using PyStan, and am getting the error “stan::variational::normal_meanfield::calc_grad: The number of dropped evaluations has reached its maximum amount (10).” When I search for this error online, I notice that it’s typically set as 100 or 500 in others’ posts. Would it be reasonable to increase this setting from 10 to 100? I also don’t see how to change this setting in the PyStan documentation for the function vb().
I believe the right thing to do is to adjust the Stan program. Please check to see if there’s some way to restrict the parameters defined so that the model can be evaluated on the domain of the parameters.
A simple example would be if you have a parameter used for standard deviation and you defined it as
real sigma;, you’d want to change that to
real<lower = 0> sigma;.