MCMC On Poisson Lognormal Model

Hi all,

I am trying to compute the posterior p(z|x) for this very simple model, and for some reason I do not understand what I am doing wrong (I am discovering this package). Generating data from this model seems to be working, MCMC sampling won’t work.

Here is my model:

def pyro_mdl(data):
#     with pyro.plate('data', len(data)):
    cell_type = pyro.sample('cell_type', dist.Categorical(dataset.probas))
    z = pyro.sample('z', pyro.distributions.MultivariateNormal(dataset.mus[cell_type], 
    exp_z = z.exp()
    x = pyro.sample('x', pyro.distributions.Poisson(rate=exp_z), obs=data)
    return x

The parameters have already been fitted (I am mostly interested in MCMC right now). When I try to perform MCMC using:

kernel = HMC(pyro_mdl, adapt_step_size=True, jit_compile=True, )
mcmc_run = MCMC(kernel, num_samples=500, warmup_steps=300).run(data=x)

I get the following error! “TypeError: only integer tensors of a single element can be converted to an index”

Do you have an idea what I have been doing wrong?

Could you paste the full stack trace? Also, note that HMC requires continuous parameters in the model, and will therefore try to integrate over all possible values in cell_type.