bambinos / kulprit Goto Github PK

View Code? Open in Web Editor NEW

29.0 29.0 4.0 4 MB

Kullback-Leibler projections for Bayesian model selection in Python

Home Page: https://kulprit.readthedocs.io

License: MIT License

Python 100.00%

bayesian-inference kullback-leibler-divergence model-selection

kulprit's People

Contributors

Stargazers

Watchers

Forkers

aloctavodia laplacekorea nsiccha enessinanparildi

kulprit's Issues

mathjax not properly rendering in docs

See https://kulprit.readthedocs.io/en/latest/examples/quick-start.html

Remove scikit-learn dependency

We use scikit-learn here

kulprit/kulprit/search/l1.py

Line 5 in dae5202

from sklearn.linear_model import lasso_path

We could implement this functionality within Kulprit

Drop support for python 3.9, add support for Python 3.12

Add support for hierarchical models

Link Kulprit documentation with Bambi's one?

To improve discoverability of Kulprit, we may want to add a section to the Bambi documentation https://bambinos.github.io/bambi/notebooks/ linking to Kulprit docs or at least one example.

Allow the ``project`` method to accept tuples

Allow a user to call ReferenceModel.project() while passing a tuple of parameter names as opposed to only either a list or an integer.

Accommodate more families

Currently only the Gaussian observation family is implemented in kulprit. We should also allow the use Binomial and Poisson observation families in the reference model.

Add support for more families

These are the families supported in Bambi https://bambinos.github.io/bambi/notebooks/getting_started.html#families we should aim to have parity.

Kulprit currently supports:

Gaussian
Poisson
Binomial
Bernoulli

UserWarning: Incompatible model and inference data.

Hi,
Beginner here. I tried to reduce my model to as few variables as possible, and follow the Quickstart, but still getting this error. Not sure how to debug.

ppi = kpt.ProjectionPredictive(model, idata)

Repro is here: https://github.com/lucabol/divs/blob/master/Analysis.ipynb

Full error below

UserWarning                               Traceback (most recent call last)
Cell In[6], [line 3](vscode-notebook-cell:?execution_count=6&line=3)
      [1](vscode-notebook-cell:?execution_count=6&line=1) #idata.to_netcdf("kptidata1.nc")
      [2](vscode-notebook-cell:?execution_count=6&line=2) #idata = az.from_netcdf("kptidata1.nc")
----> [3](vscode-notebook-cell:?execution_count=6&line=3) ppi = kpt.ProjectionPredictive(model, idata)
      [4](vscode-notebook-cell:?execution_count=6&line=4) ppi.search()

File ~/miniconda3/envs/bap3/lib/python3.11/site-packages/kulprit/reference.py:62, in ProjectionPredictive.__init__(self, model, idata)
     [60](https://vscode-remote+wsl-002bubuntu.vscode-resource.vscode-cdn.net/home/lucabol/dev/divs/~/miniconda3/envs/bap3/lib/python3.11/site-packages/kulprit/reference.py:60) # test compatibility between model and idata
     [61](https://vscode-remote+wsl-002bubuntu.vscode-resource.vscode-cdn.net/home/lucabol/dev/divs/~/miniconda3/envs/bap3/lib/python3.11/site-packages/kulprit/reference.py:61) if not check_model_idata_compatability(model=model, idata=idata):
---> [62](https://vscode-remote+wsl-002bubuntu.vscode-resource.vscode-cdn.net/home/lucabol/dev/divs/~/miniconda3/envs/bap3/lib/python3.11/site-packages/kulprit/reference.py:62)     raise UserWarning("Incompatible model and inference data.")
     [64](https://vscode-remote+wsl-002bubuntu.vscode-resource.vscode-cdn.net/home/lucabol/dev/divs/~/miniconda3/envs/bap3/lib/python3.11/site-packages/kulprit/reference.py:64) # log reference model and inference data
     [65](https://vscode-remote+wsl-002bubuntu.vscode-resource.vscode-cdn.net/home/lucabol/dev/divs/~/miniconda3/envs/bap3/lib/python3.11/site-packages/kulprit/reference.py:65) self.model = model

UserWarning: Incompatible model and inference data.

Add tutorials

Add two tutorials to the docs:

a general kulprit workflow (similar to the main projpred vignette);
an implementation of a discrete observational family using both Poisson and Binomial families.

Improve documentation

We could use the body-fat example from this paper https://arxiv.org/abs/2306.15581

I don't thing this is a priority at the moment. But at some point in the future, making this change will be beneficial. The two most obvious benefits will be removing one dependency, as aesara is already a Bambi dependency. And avoid duplicating code from bambi.

We could use this issue to track the functions that we actually need tom PyTorch that are currently absent from aesara. And then decided if those should be added upstream or we can add it as part of kulprit. I think we just require a couple of functions or so.