Thank you for your patience,
- Precompiling the random samplers with the pymc
compile_pymcutility led to similar performance in the simulator as with numpy - I moved to linux and out of jupyter notebook to rule out issues with windows or notebook
- Now with the working progress bar I noticed that one or more of the chains seems to get stuck
- Tightening the priors around hyperparameter true values (with simulated observed data) and loosening epsilon for smc let me get my first real sample for this model
I had 100 draws (/stage?) with 8 chains (and 8 cores) in this computation, which took 25 minutes to complete. All of my cpu cores are being fully utilized. For a simple comparison, if we measure roughly 0.5 seconds per draw and, hypothetically, have 10 stages for the beta evolution of SMC (assuming each stage has 100 draws as well), it would make 0.5x100x10 = 500 seconds? Sure, as you mentioned, there are more things being calculated under the hood, but should it be this big discrepancy from my rough estimation?