Fritz Obermeyer

Hi @jejjohnson, the new library is https://github.com/facebookincubator/flowtorch and is being developed by @stefanwebb @feynmanliang at Facebook. Perhaps they can comment on open source timeline and invitation-only access for early adopters.

@jejjohnson I know of no plans to port Pyro normalizing flows to NumPyro, but contributions are always welcome 😄

As a quick workaround, you could set cuda as default before creating your `spline_coupling` object ```py torch.set_default_tensor_type(torch.cuda.FloatTensor) spline_transform = T.spline_coupling(...) ``` Another not-ideal workaround is to save-and-load with [map_location](https://pytorch.org/docs/stable/generated/torch.load.html): ```py...

Hi @gshartnett, it's difficult to diagnose where the stray cpu tensor is coming from without actually diving into a debugger. I'd recommend running under pdb and inspecting each tensor's device....

👍 Hi @tillahoffmann, I think it's a great idea for Pyro and NumPyro to offer multiple parameterizations of distributions for purposes of numerical stability and interpretability. As you note we...

I love the idea of a debugging-focused tutorial. And I suspect core Pyro devs are the _worst_ people to write such a tutorial since it's been so long since we...

cc @fehiepsi

Hi @eromoe, can you try inverting your mask: ```diff - mask = torch.isnan(test_data) + mask = ~torch.isnan(test_data) ``` In Pyro's mask semantics, True means observed and False means missing.

@eromoe could you paste a little of your model code that triggers the above exception? Ideally could you provide a minimal model and fake data that triggers the error? It's...

Hi @eromoe, I'll try to debug it this weekend. Thanks for providing a reproducible example!