Discrete_markov_chain.ipynb

[NicoGreggio]

Hi everyone!

I am trying to run the discrete_markov_chain.ipynb example but I get two errors in two different parts of the code, the first when it does a demonstration of the API, in the seventh part ie:

with pm.Model() as model: x0 = pm.Categorical.dist(np.ones(3) / 3, size=(100,)) P = pm.Dirichlet("P", a=[1, 1, 1], size=(3,)) discrete_mc = DiscreteMarkovChain("MarkovChain", P=P, init_dist=x0, observed=chains) idata = pm.sample()

With the following error: ValueError: TensorType broadcastable/shape must be a boolean, integer or None, got <class 'numpy.int64'> 1

In part thirteen, the part where you create the actual MS-AR model, ie:

with hmm:

You have to assign BinaryMetropolis by hand, the default is Metropolis and it breaks.

# BinaryGibbsMetropois seems to do much better than BinaryMetropolis, but also much slower.
idata = pm.sample(
    step=[
        pm.BinaryGibbsMetropolis([hidden_states]),
        pm.NUTS([P, state_mus, ar_coefs, sigma], target_accept=0.99),
    ],
    drawings=5_000,
)

With the following error: ValueError: Random variables detected in logp graph: {bernoulli_rv{0, (0,), int64, False}.out}. This can happen when the DensityDist logp or Interval transform functions refer to nonlocal variables, or when not all rvs have a corresponding value variable.

[ricardoV94]

ValueError: TensorType broadcastable/shape must be a boolean, integer or None, got <class 'numpy.int64'> 1

For that one just update pymc and pytensor to the latest versions.

The other errors are more serious, will have to check where they are coming from.

[ricardoV94]

Found the issue, not too bad, just a change in how Shape works at the PyTensor level. Will push a fix soon