User experience: Python vs R, PyMC vs Stan, PyTensor vs JAX

[Duc_Quang_Nguyen]

@bob-carpenter

Thanks—this is really great. I’d love to hear from more users like this.

This is the common refrain I’ve heard from PyMC users—they really really don’t want to learn another language on top of Python.

It could be force of habits as I was taught C++ and Matlab during my college’s days.
Thanks a lot for providing the background knowledge about Stan (and Bayesian methods). Very good learning materials!

@bob-carpenter

Could you elaborate on what the helper functions do?

I like the ArviZ library in general. Within the PyMC, there are two functions:

1. pm.Potential . Handy function to impose constraints. I used this function to weight my data points. I need to do evidence synthesis from multiple sources, so being able to weight data is great.
2. pm.Truncated. When building model, I need to look at a particular regions of my variables for various reasons (e.g., experimenting with model structures/ functions, sourcing for priors from literature or sometimes, just to get some results because the model is not fitting properly). I do made a point to remove all these Truncated distributions in the final model.

Huge difference between MCMC posterior samples and actual data - help needed!

One thing I could pick up from your posteriors is that the range of your λ values are very large for this type of ODE (between 0.1 and 1). Is that your force of infection term? You could try to use pymc.TruncatedNormal or other truncated distributions to get a sense of where your priors on them should be (and also to , artificially, reduce your variance). After you have a good set of priors, you might consider pymc.Potential to put weights around your observed values, and remove all the Truncat…