I am trying to infer parameters of a ODE describing some concentration of chemical components. I can measure these concentrations and I know they cannot be negative. For this, I want to use TruncatedNormal distribution over the observations.
numpyro.sample("Y", dist.TruncatedNormal(np.array(concentrations), sigma, low=0), obs=Y)
But when I do that, I get Inf in the mean prediction. I thought there could be a 1/0 somewhere, so I tried to set low=10^-3
but it did not work.
If instead, I use the Normal distribution I do not have this problem.
What could cause this issue?