Theano vs. Aesara

exoplanet is built on top of the excellent PyMC3 library and it uses some of the more advanced features of PyMC3 to enable the complex models required for exoplanet data analysis. As a user of exoplanet you will become familiar with some of these details and, in particular, you will find yourself interacting with PyMC3’s backend directly. When it was originally designed, PyMC3 was built on top of a library called Theano, but the development of that library has since been taken over by the PyMC developers and it was renamed to Aesara (there are more details, and you can read about those in this blog post from the PyMC developers). Theano and Aesara are not completely compatible so this transition has been a little painful for downstream users, like you and the developers of exoplanet. In the long run, we will switch over to using Aesara directly instead of Theano, but in the meantime, since some users will have Theano installed and others will have Aesara installed, we have developed a small library that can help smooth this transition: aesara-theano-fallback. This will be installed automatically when you install exoplanet, but you can also install it directly using:

python -m pip install aesara-theano-fallback

Then, you can use this with syntax like the following:

import aesara_theano_fallback.tensor as tt


from aesara_theano_fallback import aesara as theano

In both of these examples, the library will attempt to provide a consistent interface regardless of which backend is installed. That being said, this has only been designed to support the features used by exoplanet and starry, so your milage might vary. Check out aesara-theano-fallback on GitHub for more information or to open issues.