funsor
funsor copied to clipboard
pyro-ppl
Functional tensors for probabilistic programming
(Separating this from #593 PR) This proposes the following logic for the Delta Integrate pattern: - If `reduced_var` is in `integrand.inputs` then apply substitution to `delta` and `integrand`: ```py delta...
It would be great to have reduction rules for uniform likelihoods wrt Gaussian latent variables: - [ ] `(MyContinuousDist + Uniform).reduce(logaddexp)` implemented via `.cdf()` - [ ] `(Gaussian + UniformPolygon).reduce(logaddexp)`...
Provenance tracking implementation in funsor. Here the provenance is the set of `(name, point)` tuples of RV samples that were in the history of the computations of the tracked term....
Importance sampling is represented by an `Importance` funsor. 1. Signature - `Importance(model, guide, sampled_vars)`. 2. When `guide` is a `Delta` it eagerly evaluates to `guide + model - guide`. 3....
I came across this cute example of nested semiring dynamic programming in the context of Bayesian optimization. [Thompson sampling](https://en.wikipedia.org/wiki/Thompson_sampling) first performs Bayesian regression, fitting a posterior `p(θ|Xs,ys)` over parameters `θ`...
Since Funsors are composable in clearly defined ways, could we do model structure search among possible funsors constrained only by type signature (i.e. shape signature)? Could we add an [automated...
1. Could we express ODEs in Funsor and then solve them via (approximate) numerical integration methods? 2. Could we express PDEs in Funsor and then solve them via (approximate) finite...
Addresses https://github.com/pyro-ppl/pyro/pull/2929 See [design doc](https://docs.google.com/document/d/1MVfG0po7M5UB2mNM2In6iKKq_OYNpa_I-FnJncUgXPI) This issue tracks changes needed to efficiently perform variable elimination in Gaussian graphical models with plates. While `funsor.sum_product.sum_product()` is a partial solution, we'd like to...
Consider the example using https://github.com/pyro-ppl/pyro/pull/2946 ```sh FUNSOR_PROFILE=50 python -m tests.infer.autoguide.test_gaussian --no-jit -n 2 -s 800 ``` What makes this slow? Can we implement better profiling tools to diagnose this?