NumPyroConverter makes too many assumptions about the model

The following code is not always runnable:

https://github.com/arviz-devs/arviz/blob/7ebedd2d1494f07513bc7b9ea21962bc11399a94/arviz/data/io_numpyro.py#L126-L133

This will fail for a model involving ImproperUniform, as sampling for it is not defined.

Would it be possible to add a keyword argument that specifies whether this automatic extraction of observations should be attempt or not, e.g. extract_observations or something? At the moment it’s stopping me from being able to use arviz with my numpyro-model (unless I just comment out the lines above) 😕

[torfjelde]

Sounds reasonable to me. Do you want to submit a PR? 😃

Sure 👍

I think that for coherence with those converters, the argument should be oberved_data, then for what it should take, I think a boolean to keep or disable the current behaviour makes sense, not sure if taking a dict of “extra” variables to add or override in the observed_data would make sense too.

That sounds reasonable 👍

In my opinion, those extra log probabilities play the role of log-likelihoods.

This is not true in general, e.g. in the model I’m working with we’re using ImproperUniform together with a factor in the prior to make the implementation more efficient.

[fehiepsi]

we’re using ImproperUniform together with a factor in the prior to make the implementation more efficient.

Oh, that’s a good point! I forgot that factor can be used this way. 😉