"ValueError: Bad initial energy: inf. The model might be misspecified"

Questions
1

I am modelling 2 sets of parameters, c (4 x 4) and theta (1 x 4), each having a related likelihood function like the following. However when I sample

model = pm.Model()
with model:

    # Priors for unknown model parameters
    hyper_mu = pm.Normal('hyper_mu', mu=0, sd=10)
    hyper_sd = pm.Gamma('hyper_sd', alpha=0.01, beta=0.001)
    c = pm.Normal('c', mu=hyper_mu, sd=hyper_sd, shape=(4, 4))  # We want this

    # Prior for nuisance parameter
    theta0 = np.ones(4)
    theta = pm.Dirichlet('theta', theta0, shape=(1, 4))

   
    # Likelihoods (sampling distribution) of observations
    p = sigmoid(c)
    L1 = pm.Binomial('L1', n=N_data, p=p, observed=observed_successes)
    
    p2 = pm.math.sum(L1 * theta, axis=1)  # SHAPE: (4,)
    L2 = pm.Binomial('L2', n=N_data2, p=p2, observed=successes2)

When I do model.check_test_point(), I get:

> L2                        -inf
> L1                        -7855.870000
> c                         -51.540000
> hyper_mu                   -3.220000
> hyper_sd_log__             -4.660000
> theta_stickbreaking__      -3.750000
> Name: Log-probability of test_point, dtype: float64

Which seems to suggest my L2 initialisation is bad, but I’m not sure how to correct this… I am certain that the likelihood is given by sum(L1 * theta, axis=1) and I am giving the correct data (N_data2, successes2) in the right shape (4,).

2

Usually this is your parameter goes out of the support, you can check whether p2.tag.test_value are all within 0 and 1.

3

Many thanks - this helped me understand that I was being extremely silly, I had to use p2 = pm.math.sum(p * theta, axis=1) not p2 = pm.math.sum(L1 * theta, axis=1), I thought L1 was like normalised probability similar to p. I will go back to basic documentation! Sorry for the hassle.