Skip to content

ArviZ 2021 roadmap

Jump to bottom
Alexandre ANDORRA edited this page Dec 22, 2020 · 17 revisions

Roadmap topics for 2021

TFP

  • Clean up TFP interface for current version and TensorFlow 2
  • Provide a "default trace function" that will extract the metrics arviz expects
  • Provide example of producing a fully-featured InferenceData object

Plots

  1. half-eye, dots, see ggdist https://mjskay.github.io/ggdist/
    • Better than violin, smaller, and there are better variations
  2. Dot plots
  • Useful if people need to infer talk probabilities
  1. Calibration plot for classification, see. e.g. https://avehtari.github.io/modelselection/diabetes.html
    • Multiclass would be like a pair plots
  2. ecdf / ecdf-difference with correct envelopes (more info soon in hopefully Jan)
  • loo-pit has further issues with ecdf
  • Can be used in convergence diagnostics rater than rank plot

  1. Switch to compact=True by default in our plots. Indeed, plotting more than ~10 parameters separately easily becomes cumbersome. Issue to track this goal here.

  2. Add a helper function to easily stack the chain and draws dimensions into a sample dimension, instead of having to do idata.posterior.stack(sample=("chain", "draw")). Often useful when you don't care about which chain a draw is coming from. Issue to track this goal here.

  3. Actually deprecate the use of from_pymc3 without the model. The FutureWarning has been there for a while now. Lots of people are still using the old behavior, so it's time to nudge them by doing what we said we'd do. Issue to track this goal here.

Proj pred R Package

  • Determine where this would go in python package world

Lots of chain plots

  • TFP is creating more chains than samples in some cases
  • Could help with simulation based calibration
  • Rhat computation changes based on short chains versus long chains

Refitting

  • reloo + iwmm? (Importance weigted moment matching in R Loo package and supported by BRMS)
    • Can improve results compared to psis loo with less computation time, doesn't always work
    • Can save time for people, but requires that were using the model again
    • https://arxiv.org/abs/1906.08850

Julia

  • Getting converter for gen.jl
  • Nice to have: Patch in julia's package as a backend
    • Python needs functionality to patch in backend

Generic Plotting Backend

  • Let people bring their own backend and let arviz do all the hard math stuff for them

DEI Outreach

  • Make sure we do a couple DEI
    • Use NumFOCUS money

Exploratory Analysis Repo

  • Held up on paying people to finish that. Otherwise its a big opportunity cost

Inference Data to and from R

  • A way to transfer 1 to 1 mapping from ArviZ to posterior and back
  • Need to discuss with posterior devs to see what the best way would be to do this
  • https://mc-stan.org/posterior/

Public API Stability

  • Duplicity of plot_dist and plot_kde
  • plot_hdi draws and samples conversion
  • Plotting API is not good, input and output still a mess
    • Inconsistency in changing things in plots
    • Are input arguments are all the same

Social

Funds

  • NASA Roses Grant
  • Can come up with more precise
  • Could pay for developers and DEI

ArviZCon? ArviZ social event?

  • Assume its going to be Finland

Google Summer of Code

  • Definitely

Google Summer of Docs

  • Will for this one

Onboard contributors

  • 2 or 3 new insular
    • Another jl dev if we do generic plotting backend
Clone this wiki locally