Remove `pystan` backend, move to `cmdstanpy`

[tcuongd]

Context

Pystan 3+ makes some breaking changes to Prophet python package:

• MAP estimation is removed.
• Windows support is removed.

We've been continuing to use Pystan 2.19.1.1, but this is only supported up to Python 3.8 and no longer receives development. Many of our issues also relate to pystan issues, which will become harder to debug over time and may need hacky patches.

The current installation process with pystan is as follows:

Mac OS / Linux

• End users usually use prophet from PyPi.
• End user must install pystan=2.19.1.1 first, as it is required for building the wheel
• End user installs prophet, which includes compiling the pystan model on their machine.

Windows

• End users usually use prophet from conda-forge.

Proposal

• Use cmdstan + cmdstanpy to build binary distributions for Pypi (using Github Actions) and conda-forge, across all platforms.
  • Packaging pre-compiled stan models per platform + python version means the end-user does not need to perform the build process on their machine. They won't even need their own cmdstan installation since we will package the necessary executables into Prophet itself.
  • Only the lightweight wrapper library cmdstanpy is required as a dependency.
• Remove the dependency on pystan altogether.

This should not change how end-users typically interact with the Prophet package. The general flow of

m = Prophet()
m.fit(df)
future = m.make_future_dataframe()
m.predict(future)

will remain unchanged, and all other diagnostic functions will work as normal.

It's just that behind the scenes, the model object will now be a cmdstanpy object rather than a pystan one. The argument stan_backend='PYSTAN' will also be deprecated. This will affect end users who retrieve the actual pystan object and work with it directly, but we expect this to be the minority.

We have already refactored the setup.py script to accomodate this new process (it currently supports both pystan 2.19.1.1 and cmdstanpy for backwards compatibility).
We have also built and tested PyPi wheels for macOS and Linux for Python versions 3.6-3.8. If you follow the instructions here, you can test these binary distributions on your machine.

Work required

For the next minor release (prophet 1.1.0), we should consider the following:

• Update existing conda build scripts to use the cmdstan + cmdstanpy backend.
• Remove the pystan model compilation from the build process and remove pystan as a Prophet dependency.
• (Optional) Add support for a Windows wheel on PyPi.

[mitzimorris]

excited to be the Python backend!

CmdStanPy pr stan-dev/cmdstanpy#461 changed behavoir of sampler so that CmdStanModel method sample argument show_progress defaults to True, previously default was False.

also, we're hoping that next release of CmdStanPy is going to be the 1.0 release.

[mitzimorris]

FYI - CmdStanPy PR stan-dev/cmdstanpy#460 is removing deprecated functions.
I checked class IStanBackend - this won't affect it.

[mitzimorris]

FYI, CmdStanPy 1.0.0rc1 is up. Methods, arguments, and properties which were previously deprecated have been removed. The addition of progress bars to the sample method and the cmdstan_install methods will change the console output behavoir; these progress bars use the tqdm library.

More information is available here: https://discourse.mc-stan.org/t/cmdstanpy-1-0-0rc1/24973

[tcuongd]

Thanks heaps @mitzimorris! Sorry I haven't had the time to take a look. Do you have any examples of other packages using cmdstanpy on conda btw? It'll save me some time -- if not, all good.

Would be keen to know when the 1.0 release goes on PyPI as well! Given I've been slow on this I think we might as well wait for that and refactor the setup script.

[mitzimorris]

answer from Brian who put together the conda package:
stan-dev/cmdstanpy#503 (comment)

[akosfurton]

@tcuongd and @mitzimorris , FYI the 1.0 release of cmdstanpy is available on pypi

https://pypi.org/project/cmdstanpy/

related discussion: stan-dev/cmdstanpy#475

[WardBrian]

Are you looking to compile the .stan files as part of the conda build? Or just depend on cmdstanpy and actually compile the executables at install or run time?

I'm happy to help with this if I can provide value

[akosfurton]

@WardBrian , yes that would be the idea. Would love the help since this issue is blocking a new release to pypi

Packaging pre-compiled stan models per platform + python version means the end-user does not need to perform the build process on their machine. They won't even need their own cmdstan installation since we will package the necessary executables into Prophet itself.