arviz-devs / arviz

Exploratory analysis of Bayesian models with Python
https://python.arviz.org
Apache License 2.0
1.6k stars 401 forks source link

R and Julia #499

Open ahartikainen opened 5 years ago

ahartikainen commented 5 years ago

We should have some functionality to transform results gathered with R or Julia to our InferenceData format (netCDF4).

These could be native (written in R and Julia) modules.

sethaxen commented 5 years ago

I'm currently implementing a Julia wrapper for ArviZ called Arviz.jl. It includes a from_mcmcchains function to convert an MCMCChains.Chains object into an InferenceData. That's the output type returned by a number of PPLs in Julia, including Turing.jl. It still needs tests, documentation, and more support for ArviZ’s API, but I’d appreciate any feedback.

sethaxen commented 5 years ago

Julia has a missing value similar to a nan that explicitly annotates a value that is missing. Variables in infinite mixture models in Turing sometimes take on this value. Do other packages you provide converters for have something similar? Should these kinds of values just be represented with nans in arviz?

sethaxen commented 5 years ago

Also, is there a place where all of the recommended conventions for sample_stats names are documented?

ahartikainen commented 5 years ago

Hi, I think nan is currently used for missing data, but I don't think we have tested it.

I'm not sure if we have updated list.

Maybe we could create a wiki for the common names (divergent etc)

mike-lawrence commented 3 years ago

FYI I'll be working shortly on translating cmdstan CSVs to InferenceData (netcdf4) from R. I'll be posting my work here as part of a larger project, but will propose a PR when I have something arviz users might find useful.