Hi @fritzo
I added the constraints. I forgot about those.
However even with the constraint the value becomes nan after the first iteration. Also, for the first iteration, my loss is infinite after I take the SVI step. If I make the learning rate smaller, say 0.0000000000001, then the SVI goes for two iterations and then one of the other values becomes nan (in this case one of the scale values for beta_{i}).
For the censoring logic, I’m using scipy’s CDF for the Gamma distribution which is why the code there is messier