SharedPosterior

class pymc_marketing.pytensor_utils.SharedPosterior

Posterior draws held in PyTensor shared variables.

extract_response_distribution() conditions a graph on posterior draws by substituting them as constants, which welds every graph – and every function compiled from it – to one posterior. Passing a SharedPosterior instead binds the draws through shared variables, so set_posterior() can point an already compiled function at a new set of draws without extracting or compiling again. That is what makes a loop over many posteriors (a resampled or updated posterior per iteration) affordable: the compile happens once.

One instance can back several graphs. A variable is created the first time a graph needs it and reused by every later extraction that passes the same instance, so a single set_posterior() call rebinds all of them at once. A later extraction must therefore be conditioned on the draws the instance already holds; handing it a different posterior raises, since the graph would otherwise silently read the bound draws.

The contract for a rebind: every bound variable is present with the dtype it was bound with; every non-sample dimension keeps its name and its set of labels, compared by value (so a reordered channel axis is realigned, and a datetime axis may change storage unit); a dimension bound without labels must arrive without labels; the number of draws is free. Anything else is refused, and a rebind is atomic: every variable is validated and laid out before any is written.

Examples

shared_posterior = SharedPosterior()
graph = extract_response_distribution(
    model, idata, "channel_contribution", shared_posterior=shared_posterior
)
fn = function([budgets], graph)  # compiled once

fn(x)  # under ``idata``
shared_posterior.set_posterior(other_idata)
fn(x)  # under ``other_idata``, no recompile

Methods

SharedPosterior.__init__()
SharedPosterior.set_posterior(idata)	Rebind every variable to the draws in idata.

Attributes

SharedPosterior.dims	The dimension order each variable is stored in, sample first.
SharedPosterior.variables	The shared variables, keyed by posterior variable name.