Using Custom blackbox likelihood and prior functions

1

Hi,

I am looking to do a simple MCMC run to sample from a posterior distribution using 1000 samples extracted from a previous analysis as starting points. My posterior function (likelihood and posterior) uses numpy and scipy as well as if statements and since I also rely on scipy.root simply rewriting the function to work with aesara is not possible. I do of course realize that this will decrease performance, but I would like to implement this regardless. However, everything I have tried still throws errors such as “object of type ‘TensorVariable’ has no len()”, which indicates that my variables are still interpreted as tensor variables. Here is a short version of my current code:

with pm.Model() as model:
# Define parameters
x = pm.Flat(“x”, shape=(run.n_dim,))

# Add log-likelihood and log-prior as potentials
pm.Potential("log_likelihood", run.log_like(x))
pm.Potential("log_prior", prior.logpdf(x))  # your log prior function

# Use Metropolis sampler
step = pm.Metropolis()

# Set up starting points from predefined 1000 samples
start_points = [{"x": pt} for pt in samples]

# Sample
trace = pm.sample(draws=1000, tune=500, step=step, chains=1000, cores=your_cores_count,
                  start=start_points, progressbar=True, discard_tuned_samples=True)

I have tried to wrap the functions using @as_op(itypes=[at.dvector], otypes=[at.dscalar]) and nothing seems to work.

Is what I’m looking for even possible with pyMC?

2

The example notebook walks slowly over the steps to create your own op and test the implementation: Using a “black box” likelihood function — PyMC example gallery