I am trying to learn 12 parameters over a simplex. The data to fit is a matrix P that is (K, 12) in shape and the observation vector (K, 1)-shaped. I am using this model:
with pm.Model() as model: f = pm.Dirichlet("f", np.ones(12)) n_ = pm.Binomial("n", n=2, p=pm.math.dot(P, f), observed=n) idata = pm.sample()
K is large, like O(1e5) rows. But still, I think this is a relatively easy problem - and MAP can be computed instantaneously as well. However, sampling from the above model takes hours. Is this normal?