Any updates on PyMC4 and the usage with tensorflow_probability?


#1

I haven’t seen anyone on PyMC4 since the development from the Google Summer of Code with the schools dataset example. Is there any updates on the API? Does anyone know if there will be a functional approach like Keras?
Nox VidMate


#2

Yes, there have been major changes done to the PyMC4 API after the development meeting at London last year. You can check everything at the github repo. It is still pre-release and the API is very much open for debate. Currently, the model follows a functional approach, with some context managers to control whether use the model for forward sampling (like pymc3.sample_prior_predictive and pymc3.sample_posterior_predictive), or for inference (like pymc3.sample).

The distributions inherit from TensorFlow Probability, but we chose to keep a more consistent API with pymc3 (i.e. distribution parameter names). Again, to get up to date with the status of PyMC4, check the github repo.