Originated as a forum post

Let's add a tutorial demonstrating how to serve trained models in C++ using torch.jit.save() and torch::jit::load(). I believe there is some subtlety requiring a torch.jit.trace() to wrap poutine.trace and poutine.replay logic for SVI mod

Description of your problem

Most distributions support multiple parameterizations that are hard to grasp from the little description that we have. In some cases the scipy.stats documentation is not helping either.

Proposal / Ideas for improvement

• state what the default parameterization would be
• show examples with both (equal) parameterizations
• link to the scipy.stats distrib

I think there is a little typo which gives the following error:
TypeError: init() got an unexpected keyword argument 'mu'

I checked the "mixture_gaussian_gibbs.py" example and I suggest the following suggestion
params = {'loc': tf.zeros(5), 'scale': tf.ones(5)}

https://github.com/blei-lab/edward/blob/352596c3bc77e1cb5c01f77ac3283a53c74dbf21/edward/models/param_mixture.py#L36

SeparableConv enables a much greater range of low power, low latency models. MobileNet and other low cost high performance networks generally utilizes SeparableConv for it's efficiency. I am building a real probabilistic inference version of a MobileNet which uses SeparableConv. I am using regular tfp convolution ops for now for initial validation but it is much less efficient and will not be fit

Ankit Shah and I are trying to use Gen to support a project and would love the addition of a dirichlet distribution

I don't know what the plans are for this software, but if there is a reasonable chance that it will require bug fixes or new features in the future, I would recommend that someone experienced with the system spend maybe 1-3 hours on basic developer documentation - just enough to orient a new person to the code.

If no more work will ever be done, this issue is moot and should be closed as a 'won

Summary:

Documentation of optimizing does not appear accurate, specifically the init parameter

Description:

The description of the init parameter of optimizing is identical to that of sampling, with reference to chains, which optimizing does not use (as far as I understand).

Behavior does not seem to correlate with choices of init either. For example, initializing

The following pkgdown R library allows generating a doc site based on GitHub repo information (e.g. the README.md, developers etc). Wondering whether we can do something similar with Turing sub-modules, particularly build a page from the README.md file and release notes.

https://pkgdown.r-lib.org/

(Assuming it's part of the public API.)

Refer to the discussion in https://forum.pyro.ai/t/hmm-like-model-with-sequences-of-different-lengths/1507.

I've made a small change to the documentation, then tried to push them live (via the handy-dandy script) and it failed miserably. Tried to fix my installation of mkdocs, to no avail. @zaxtax can you give it a try?

The current example on MDN from Edward tutorials needs small modifications to run on edward2. Documentation covering these modifications will be appreciated.

• add note that carat is actually bitwise XOR so the things generated from the first box shouldn't actually be run (they later get turned into the right thing in runify

Remove examples that are not part of the book. Migrate to webppl repository, forest, or specific research repositories as appropriate.

https://github.com/probcomp/metaprob/blob/c1a01ea02c51ea0f8fedebd45954914d7645bde9/src/metaprob/distributions.clj#L55-L60

In the scorer, argument x should be checked to lie in the interval [a,b].

Although the collection of base Kernels in kernels.jl is significantly broader than it used to be, there's still quite a number of kernels that are listed in, for example, the GPML manual that aren't implemented here. A starting list of un-implemented kernels

This issues tracks changes needed to port the BART forecasting example to funsor.

To reproduce

To run a recorded version of the example, try:

cd funsor
pytest -vs test/examples/test_bart.py --runxfail

To run the original example (requires downloading data):

git clone git@git