"Injecting" values into each MCMC iteration

Questions v5
2

What I tried so far:

  • Create a CustomDist for the y_obs with a custom logp function that draws from a python generator.
  • Create a similar logp function that goes into the pm.Potential for the y_obs.
  • Create a CustomDist for w which has a generator for both logp and random.

But these seem to use only a handful (1-10) draws from the weights matrix (I keep track with an internal counter), instead of throughout the sampling.
However, my knowledge in the inner-works of pymc is very limited, I don’t always understand what I try :slight_smile: