Fritz Obermeyer

Oh thanks @fehiepsi I'll try to release soon

@RylanSchaeffer there is no tutorial on migrating distributions. To migrate a distribution from Pyro to PyTorch you'd need to - add code to torch/distributions/dirichlet_multinomial.py or torch/distributions/conjugate.py - fix docstrings to...

More generally I think bags-of-samples are a more flexible data structure and could be used throughout Pyro & NumPyro. Marco & Vikash said they found bags-of-samples to be the right...

> Could you suggest a way to implement this? Sorry I missed this. In many places we use sets of samples encoded as a batched tensor of shape `(num_samples,) +...

Hi @fonnesbeck, can you provide more details to the inference you are performing? It sounds like you might be using MCMC? Assuming you are using MCMC: One issue with saving...

@fonnesbeck thanks, interesting, I wasn't aware that SVI was discarding deterministic sites. Could you provide a little code snippet to show how what is going wrong? Is it specifically using...

Thanks @fonnesbeck, that makes sense. No need for an example. I think we can support deterministic sites via a new `.predict_quantiles()` method. ```py class AutoGuide(PyroModule): ... @torch.no_grad() def predict_quantiles(self, quantiles,...

See also https://github.com/pyro-ppl/pyro/issues/1502

Nice trick! I guess the grad trick will work for reparametrized continuous distributions, but not discrete distributions.

@vitkl @jamestwebber can you elaborate on the task you're trying to accomplish? From @jamestwebber's comment > I find myself in a similar situation, trying to concatenate posterior samples it sounds...