I’m afraid it’s not (at least with my current knowledge of math).
Moreover I was looking at this pull request in which gamma CDF was not implemented due to missing a incomplete gamma function implementation.
There are two possible way to handle this:
- Wrap mpmath package incomplete gamma/maijer G function implementation in a theano op, but I expect this to be very slow (functions don’t support vectorization/broadcasting natively)
- Use lognormal distribution instead of gamma
I guess the second option is the most feasible one.