I want to have a fair comparison between different mcmc methods, particularly, I want to compare the sampling performance of my method with NUTS. However, I think the sampling function in pymc3 has lots of settings/tuning which improve the performance of the sampling algorithms beyond the original methods.
So, I was wondering if anyone could tell me how to disable those tunings. In short, I just want to have a simple NUTS algorithm (by NUTS, I mean Algorithm 6: Efficient No-U-Turn Sampler with Dual Averaging)
Thanks
This might be difficult -- both PyMC3 and Stan (some of whose maintainers wrote the NUTS paper) have incorporated new best practices and improvements. You might clone it from github and check out an early implementation.
This commit has a NUTS implementation that follows the notation from the paper pretty closely. It has since been refactored to use more descriptive variable names along with the above improvements.