bayesian-inference

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

Summary:

It'd be nice to have a builder pattern for var contexts to make them easy to construct for testing. Something that could be used like this:

MatrixXd m(3, 2); 
...
var_context vc
  = var_context::builder()
    .matrix("a", m)
    .real("f", 2.3)
    .build();

Current Version:

v2.23.0

In SVI latent random variables sampled from the guide completely mask those sampled in the model during inference. However, nothing prevents us from specifying different shapes for such a sample site in model and guide respectively. This makes it easy to introduce confusing bugs when code in the model expects a certain shape that is different than what guide provides. This can easily happen if the

trace_to_dataframe() in PyMC3 to save traces is currently implemented in Rethinking_2 notebooks (e.g. Chp_04). But the function is planned for deprecation, with Arviz being the intended package to save traces. As per this comment by @AlexAndorra, Arviz's InferenceData format is a superior replacement to this function as it

Hi @JavierAntoran @stratisMarkou,

First of all, thanks for making all of this code available - it's been great to look through!

Im currently spending some time trying to work through the Weight Uncertainty in Neural Networks in order to implement Bayes-by-Backprop. I was struggling to understand the difference between your implementation of `Bayes-by-Bac

Hi,

is there any plan to implement the Generalized Pareto Distribution in brms (paul-buerkner/brms#110 (comment))? I am playing around with an extreme values analysis and it looks like extremes collected as Peak Over Threshold are better represented by the GPD instead of the generalized extreme value distribution, which I am so happy to see already in `b

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