Bayesian Classifier Combination Model (need help)

Questions
4

Hello @junpenglao,

Thanks for your response and advice!

The shapes of the variables within the steps above are:

  1. Dirichlet sample of class probabilities [1,]
  2. Sample of class labels t [I,]
    Hyper parameter lambda [J, J]
  3. Sample of alpha [K, J, J]
  4. Confusion matrices sample [K, J, J]
  5. Class labels generated by classifiers [I, K]

I suppose using the T.tensor.stack([]) function the code would look like this:

IBCC = pm.Model()

with IBCC:
    p_vec = pm.Dirichlet(name='p_vec', a=nu_vec)

    t = pm.Categorical(name='t', p=p_vec, shape=I)

    alpha = [pm.Exponential('alpha_{}'.format(k), lam=lambda_mat, shape=(J, J)) for k in range(K)]

    alpha = T.tensor.stack(alpha)

    pi = [[pm.Dirichlet('pi_j{}k{}'.format(j, k), a=alpha[k][j], shape=3) for j in range(J)] for k in range(K)]

    pi = T.tensor.stack(pi)

    c = [[] for k in range(K)]

    t_sample = t.random()

    for i in range(I):
    
          for k in range(K):

                c[k].append(pm.Categorical('c_k{}i{}'.format(k, i), p=pi[k][t_sample[i]], observed=c_train[k][i]))
    
    c = T.tensor.stack(c)