update to pymc 5 and pytensor

[aloctavodia]

This is the failing test:

def test_shared_variable():
    X = np.random.normal(0, 1, size=(50, 2))
    Y = np.random.normal(0, 1, size=50)

    with pm.Model() as model:
        data_X = pm.MutableData("data_X", X)
        mu = pmb.BART("mu", data_X, Y, m=2)
        sigma = pm.HalfNormal("sigma", 1)
        y = pm.Normal("y", mu, sigma, observed=Y, shape=mu.shape)
        idata = pm.sample(tune=100, draws=100, chains=2, random_seed=3415)
        ppc = pm.sample_posterior_predictive(idata)
        new_X = pm.set_data({"data_X": X[:3]})
        ppc2 = pm.sample_posterior_predictive(idata)

fails with

<<!! BUG IN FGRAPH.REPLACE OR A LISTENER !!>> <class 'TypeError'> The type of the replacement (TensorType(float64, (3,))) must be compatible with the type of the original Variable (TensorType(float64, (50,))). random_make_inplace

ValueError: shape mismatch: objects cannot be broadcast to a single shape.
Mismatch is between arg 0 with shape (50,) and arg 1 with shape (3,).

EDIT by Michael: I copied the failing test code here for reference.

[aloctavodia]

@ricardoV94 @michaelosthege Do you have a quick idea of why the shape test is failing?

[michaelosthege]

So the <<!! BUG IN FGRAPH.REPLACE OR A LISTENER !!>> sounds like the rewrite self-diagnosed that there's a bug.

Good thing it happened with such a simple MRE !

Something about the BART_rv or a rewrite that is related to it must have choked on the resizing of the data_X SharedVariable from (50,) to (3,).

My best guess is this line:

pymc-bart/pymc_bart/bart.py

Line 46 in 9aa4896

shape = self.X.shape[0].eval()

This results in a static shape and the rewrites that happened later might have relied on this incorrect static shape assumption.