Created an issue #597 in funsor. I also tried using TruncatedNormal with low=0 instead of TransformedDistribution with ExpTransform, but I get the same error: ValueError: Invalid shape: expected (), actual (1,). I have added that info the feature request as well.
Can you suggest a workaround in the meantime? As I mentioned, the sampler converges but it doesn’t produce the right values. I don’t think what I’m doing above is right.
For instance, in the model in this comment, I am setting theta_5_raw ~ normal(mu, std) and doing theta_5 = jnp.exp(theta_5_raw), whereas I actually want theta_5 to be normally distributed with the same mu and std parameters.
Same is the case with the orderd constraint for the model in this comment.
Maybe I need to use numpyro.factor() here to update the log pdf, ala updating the target variable in Stan?