Hi Everyone,
tl;dr
Drawing from a Dirichlet distribution with shape (1,N):
with pm.Model() as model:
α = pm.Dirichlet(r'α', a=5*np.ones((1, N)), shape=(1, N))
appears to work in pyMC3 < 3.9 but fails in more recent versions with the error:
Bad initial energy, check any log probabilities that are inf or -inf, nan or very small:
α_stickbreaking__ -inf
CRITICAL:pymc3:Bad initial energy, check any log probabilities that are inf or -inf, nan or very small:
α_stickbreaking__ -inf
pymc3.parallel_sampling.RemoteTraceback:
Context
I’m looking at data that appears to be Poisson distributed according to an observed rate Λ(t) (with T observations over time) and has been sharded into N partitions. I’d like to model the relative share that each partition has of the overall Poisson count.
To do that I’m running N Poisson regressions (one for each partition) with a rate given by Λ*α where α is the share of the full rate ‘owned’ by that partition. The α’s I’m drawing from a Dirichlet distribution.
Here’s a sketch of the model:
with pm.Model() as model:
Λ = pm.Data('Λ', Lambda) # Total observed rate per time step. Lambda has shape (T,1)
# Model for share of underlying Poisson distribution
α = pm.Dirichlet(r'α', a=c*np.ones((1,N)), shape=(1,N))
# Poisson rate per partition
λa = pm.Deterministic('λa', pm.math.dot(Λ,α)) # Should have shape (T,N)
# Observed counts per partition
â = pm.Data('â', a_hat) # â has shape (T,N)
a = pm.Poisson("a", mu = λa, shape=a_hat.shape, observed=â)
Having the shape of the Dirichlet being (1,N) is to ensure that it broadcasts against the batch of different observations Λ with shape (T,1). Now the above model seems to work great in versions of pyMC < 3.9 but in 3.9 and above I’m seeing the above error. I’m not sure if I understand why a version change should induce weird values in the logp and couldn’t see anything in the release notes to suggest so. Is there a better way I can be going about this?
Thanks!
