Fritz Obermeyer

Hi @robsalomone, I believe normalizing flow development has moved to https://github.com/stefanwebb/flowtorch and then moved to https://github.com/facebookincubator/flowtorch . @stefanwebb What would you recommend as the best place for current Pyro users...

> taking forever that's probably the reason we didn't add this to ci

Could you provide your torchvision version? I believe you'll need torchvision 0.12. I believe the `MNIST` example requires torchvision's dataloader, and the segfault may be due to a torch/torchvision version...

> Would it also make sense to refactor the automatic transformation to unconstrained space... Interesting, this seems similar to @fehiepsi's `uniform_reparam_transform()` in https://github.com/pyro-ppl/numpyro/pull/807. Before we do any major refactoring, I'd...

not sure how to restart in github actions, but you could merge Pyro dev branch and push, that should rerun ci

Hi @nb2838, glad to hear you're implementing this in Pyro! I can answer directions directly on your doc if you turn on commenting. Here are some answers: > The algorithm...

@nb2838 no hurry, feel free to take as long as you like. I've tagged this `help wanted` just to signal that "no core Pyro contributor plans to work on this...

@nb2838 IIUC you need the underlying randomness to agree between two elbos, but the parameters to differ? I believe you can do this in Pyro for a subset of distributions:...

@nb2838 Sounds good. I think it's safe to first implement inference with a check (the `ValueError` above) that assumes models have been manually reparametrized, and then in follow-up PRs we...

@stefanwebb is redesigning the normalizing flows interface for PyTorch-Pyro in a [new library](https://github.com/stefanwebb/simplex). We might want to wait a bit to see what new abstractions he decides on.