Probabilistic Models of Cognition, in Pyro

Overview

Work-in-progress project to translate Probabilistic Models of Cognition (github) from WebPPL to Pyro. This collection of notebooks might help audiences without much background in probabilistic programming start using Pyro, building basic knowledge of these powerful tools so readers can begin to apply them to more complex domains.

A second Pyro project more akin to DIPPL examining Pyro's implementation with something like minipyro seems a logical next step after this - though that book would diverge more from its WebPPL counterpart.

Requirements

python >= 3.6

pyro-ppl >= 0.4.0
jupyter
matplotlib
seaborn
tqdm >= 4.31.1

PyroWebPPL

We introduce a simple, minimal implementation of WebPPL operators using Pyro as a backend for key PPL functionality. Code for this implementation can be found in the pyro_webppl directory for this project.

The PyroWebPPL examples can serve as a stepping stone to learning Pyro. All readers are encouraged to explore pyro_webppl/webppl.py and pyro_webppl/infer.py to better understand Pyro. Adventurous readers may add breakpoints or print statements to the source code to examine pyro variables used (e.g. traces in MH inference) in the relatively trivial examples present in this book.

Contributing

Contributions are welcome! See chapters and TODOs below for ideas.

Progress

Chapters

• Done: 1, 2, 3, 4, 5 (physics sims missing in 2, 3 - see below)
• In Progress: 6
• Draft Only: 7, 8, 9, 10
• Not Started: 12, 13, 14, 15, 16

Example

Nothing here yet.

PyroWebPPL TODO

pyro_webppl/infer.py

• MH
• MH drift kernels
• Sequential Monte Carlo (SMC)
• 'optimize' (with SGD, Adam, ...)
• Use HashingMarginal and viz_marginals for functions that don't return torch.Tensor - needed for Ch. 10
• reflectance/illumination example in ch 5 is slow

pyro_webppl/webppl.py

• marginalize (ch 6)

• posterior_predictive.MAP() (ch 6)

• correlation (ch 6)

• viz_marginals (ch 4)

• viz_density (ch 5)

• viz_heatmap (ch 5)

• viz_scatter (ch 5) - maybe unnecessary

• unit tests

• Physics example (ch 2)

• Drawing examples (ch 7)

Javascript <-> Python

The physics examples are a bit tricky to implement natively in Python. I spent some time exploring pymunk and pygame but had trouble getting examples to run in jupyter notebooks, essential for this project.

Notebooks natively support %%javascript cells and you can pass data between python and js (example), so a particularly convenient solution would be to load the javascript libraries used in probmods directly and use these in the appropriate chapters.