Cannot convert RV to a tensor variable (Aesara)


I am trying to compute the likelihood of a hidden markov model (HMM) marginalizing over the hidden states and, in order to do that, I am following the example in How to wrap a JAX function for use in PyMC. However, I cannot convert random variables to a tensor variable using aesara.

I was able to compute the marginal HMM likelihood using JAX, but I cannot run the PyMC model that is presented in the notebook. Here is the model:

with pm.Model() as model:
    emission_signal = pm.Normal("emission_signal", 0, 1)
    emission_noise = pm.HalfNormal("emission_noise", 1)

    p_initial_state = pm.Dirichlet("p_initial_state", np.ones(3))
    logp_initial_state = at.log(p_initial_state)

    p_transition = pm.Dirichlet("p_transition", np.ones(3), size=3)
    logp_transition = at.log(p_transition)

    loglike = pm.Potential(

When I run this model, I get the following error:

NotImplementedError: Cannot convert p_initial_state ~ Dirichlet to a tensor variable.

A similar question was asked in this post and the proposed solution was to use theano instead of aesara, but I think that it is not a solution for my problem.

Is the problem related to the versions of python, aesara and PyMC?

Here are the versions that I am using:

  • Python: 3.9.12
  • pymc3: 3.11.4
  • aesara: 2.6.6
1 Like

PyMC v3 relies on theano (see here). PyMC v4 will rely on aesara (see here).