martinjankowiak
commented
2 years ago @chadrs2 hello chad, thanks for the suggestion. we're aware of this gap but have not had time to write such a tutorial. one reason for that is that the answer to "how does pyro use distributions" is a bit complicated and depends on the particular inference algorithm. for example if you use a distribution d in a model and use SVI or HMC for inference, all you really need is to implement d.log_prob: a sample method is not strictly necessary (although might be used for initialization in the HMC case). if you use d in a guide in SVI, however, then you need a sample method, and indeed you probably want a rsample method to get reparameterized gradients if doing so is possible and as described in this tutorial.
note that this design doc, while a bit old, may be a helpful place to look for generic details about pyro's distributions api (which is built on top of pytorch's).
I'm using SVI in pyro and one of the model distributions, torch nor pyro supports. I had to develop my own implementation of that distribution. Can you create some clear guide on how to develop your own distribution / how pyro uses it? This would be extremely helpful when debugging my code and improving runtime.