issues when numpyro.enable_validation(is_validate=True)

[hyperfra]

Hi I identified a strange behavior: I estimate a VectorAutoRegression model and when I use numpyro.enable_validation(is_validate=True), (which I had in the code for some versions of the model with a MultivariateTruncated) , the resuts of any simple version of the model (with MultivariateNormal residuals say) were off. Taking out numpyro.enable_validation(is_validate=True) the results look fine. Should I be worried?

best

[fehiepsi]

I think there might be some problems. When enabling the validation, we mask out-of-support samples by zero density. Not sure if it is related. Could you make reproducible code?

[hyperfra]

Ok here is an example, if you comment line 11 it returns sensible results. Otherwise it fails. The data are downloaded via pandasdmx. If you prefer I can send you a csv.

[martinjankowiak]

@hyperfra downloading zip files from unknown third parties isn't exactly a great idea. can you please instead provide e.g. a gist and create fake data within the script?

[hyperfra]

sorry, you are quite right. now the link works https://gist.github.com/hyperfra/9e8df68414b0172c852a16db82659bc6

[fehiepsi]

Thanks @hyperfra! It seems that there is something wrong with the validation logic of LKJCholesky distribution.

[hyperfra]

ok :-)