Poisson discrete latent variable

numpyro
1

Hi,

I am running a model using DiscreteHMCGibbs.
Here is a toy model with the same issue:

def test_model(args):
  ...

  sigma = numpyro.sample('sigma', dist.HalfNormal(0.2))
  with numpyro.plate("nodes", L, dim=-1):
      alpha = numpyro.sample("alpha", dist.Poisson(sigma))
      beta = numpyro.sample("beta", dist.HalfNormal(1))

  N = numpyro.sample("N", dist.Poisson(a + b) , obs=Nobs)
  return N

The issue is that the model does not recognise alpha as a discrete latent variable and instead it outputs the following error

AssertionError: Cannot detect any discrete latent variables in the model.

If I try and run the model using only NUTS or HMC I get this error:

RuntimeError: This algorithm might only work for discrete sites with enumerate support. 
But the Poisson distribution at site alpha does not have enumerate support.

So I assume it recognises the Poisson as a discrete latent variable in the model using NUTS/HMC.

I am relatively new to numpyro so apologies if the issue is something obvious, but does anyone know of any solutions?

Thanks

2

i believe that algorithm only allows for discrete latent variables with bounded support. so you would need to use a truncated poisson or similar

3

Thanks for the reply.
Is there an algorithm that allows for the use of the unbounded Poisson?