Ami Falk

I find I often have a pattern where my random variable is a nuisance variable, but some deterministic function of it is meaningful. In this case, the desired behavior is...

It looks like all the samplers create a trace on initialization, most via `initialize_model`. It should be easy to add a function to `infer.util` that takes the trace and returns...

Is there a way to add this functionality (even if only for a subset of models), or is it a limitation of numpyro?

Sorry, no updates currently @pavleb. We ended up resolving speed issues by moving over to numpyro.

It occurs to me that adding batched mapping directly to `nmap` may not be the right thing to do here. I'm thinking about this feature in the context of batching...

I wrote up a prototype of this functionality described above ^ here: https://gist.github.com/amifalk/e21059da7f0c0ecb3db8240604413998 I realized there's no benefit to allowing different batch sizes for different axes given that they're all...

Agreed w.r.t. all points emulating `jax.lax.map`. My gut says the name `pz.nx.map` is fine given that it will have a different function signature then `nmap`, but I don't have a...

Yep, just a little syntactic sugar :). When training large models, I often use a data loader to grab data from the disk and process it onto a numpy array,...

It depends on how many chains you run, whether or not you have a gpu, the amount of native python code in your model, etc., but it can often be...

`get_extra_fields` currently only allows you to extract internal metadata generated by the sampler at each step. If want to access some intermediate value of your likelihood function, you can recompute...