Simple tests of full models

[jaanli]

Unit tests should include things like linear/logistic regression or a simple deep latent gaussian model, either on MNIST or synthetic data.

This will help make sure any changes to the API recover the same ELBO across a suite of test models.

If MNIST is too computationally intense, we can create synthetic data.

[poolio]

Great idea! Having a set of synthetic datasets would also be useful for comparing techniques, kind of like what scikit-learn does for clustering: http://scikit-learn.org/stable/modules/clustering.html

[dustinvtran]

Thanks for the reference! Yep, sounds useful. @mariru is working on that in #125

[dustinvtran]

Fixed in #487