Speeding up Bioscrape Inference

[WilliamIX]

Emcee is designed to deal with parallelization - we should make use of this!

[ayush9pandey]

I will bring this up again in BE240 recitation to see if we can get help on this from someone...

[ayush9pandey]

Reviving this issue, since there has been some progress on it over the last few months. A related issue is to add GPU compatibility for inference.

@Farnazmdi (bioscrape reviewer for JOSS) suggested Numpyro as a possible option that may speed up the inference module.

@WilliamIX I believe you have worked on both of these aspects - parallelization and GPU compatibility. Do you want to add your thoughts here?

[WilliamIX]

Thanks for suggesting this @ayush9pandey and @farnazmdi. Short answer is that the bulk of Bioscrape was written before Numpyro and similar tools were commonly used and stable. I agree that such libraries provide a route for potential GPU compatibility, however they would likely require a complete overhaul of the entire library switching from compiled Cython code to just-in-time compilation methods with completely different dependencies and implementation.

At a higher level, there are a number of GPU-enabled ODE and SSA solvers that have become available in the past few years. I think these are better options for GPU-based inference approaches. Bioscrape is not aiming to be the fastest inference engine possible, but rather strike a balance between a functional set of unique features (many of which aren't readily available in many simulators, such as stochastic delays) and speed.