probabilistic-programming

Pyro's HMC and NUTS implementations are feature-complete and well-tested, but they are quite slow in models like the one in our Bayesian regression tutorial that operate on small tensors for reasons that are largely beyond our control (mostly having to do with the design and implementation of torch.autograd), which is unfortunate because these

tfd.Categorical.quantile not working

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

Having trace.format_shapes() as in Pyro would help with debugging and verifying models, especially with enumeration, as shown in the Pyro tutorial.

Is your feature request related to a problem? Please describe.
While sales forecasting, it is necessary that the model is given the input about the promotions, special events that are taken care of in the prophet model as the holiday effect. Does orbit support this feature?

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

It is hard to keep track of how subsequent changes affect the sampling performance of the library. We should add ˋpytest-benchmark` to the test suite and use related github action to run benchmarks automatically.

There are a variety of interesting optimisations that can be performed on kernels of the form

k(x, z) = w_1 * k_1(x, z) + w_2 * k_2(x, z) + ... + w_L k_L(x, z)

A naive recursive implementation in terms of the current Sum and Scaled kernels hides opportunities for parallelism in the computation of each term, and the summation over terms.

Notable examples of kernels with th

See wikipedia table for example patterns: https://en.wikipedia.org/wiki/List_of_integrals_of_Gaussian_functions