Remove testing dependency on http download for radon dataset

[utkarsh-maheshwari]

Description

Modified loo_pit tests to remove dependency on radon dataset. Using modified centered_eight instead.
Closes #1716

Checklist

• Follows official PR format
• Code style correct (follows pylint and black guidelines)
• Changes are listed in changelog

[codecov]

Codecov Report

Merging #1717 (98453d5) into main (4d9caca) will not change coverage.
The diff coverage is n/a.

❗ Current head 98453d5 differs from pull request most recent head cd99aff. Consider uploading reports for the commit cd99aff to get more accurate results

@@           Coverage Diff           @@
##             main    #1717   +/-   ##
=======================================
  Coverage   90.86%   90.86%           
=======================================
  Files         108      108           
  Lines       11821    11821           
=======================================
  Hits        10741    10741           
  Misses       1080     1080           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4d9caca...cd99aff. Read the comment docs.

[canyon289]

This looks fine to me. Im not sure why code coverage is decreasing but testing wise it seems fine. In regards to your other comment I'm not entirely sure what this is testing either but Ive only just glanced at it

[utkarsh-maheshwari]

Isn't it safe to merge it?

[OriolAbril]

The modified test is failing to trigger this else condition: https://app.codecov.io/gh/arviz-devs/arviz/compare/1717/changes#D1L1649. It's nothing to worrying but it should not be difficult to generate synthetic data that ensures this case is tested, I'll try to comment with an option doing this.

[canyon289]

The modified test is failing to trigger this else condition: https://app.codecov.io/gh/arviz-devs/arviz/compare/1717/changes#D1L1649. It's nothing to worrying but it should not be difficult to generate synthetic data that ensures this case is tested, I'll try to comment with an option doing this.

@utkarsh-maheshwari if you able to can you see if you can trigger this else case, perhaps with a parameterized test, or a separate test? After that we should be good to for a merge. I know I'm asking for extra work but it certainly will be much appreciated

[OriolAbril]

This test checked two things: 1) that loo pit works in all possible cases and that it always returns values between 0 and 1 and 2) that loo_pit works when there are multiple likelihood variables present. The "works in all cases" roughly means having observed data (y) in multiple quantiles of the posterior predictive (y') samples and having both y>y'_i and y<y'_i for all i. The multiple likelihoods test translates into having two variables in log_likelihood group, the decoy filled with zeros is great.

Here is one approach for a synthetic dataset that does all this:

rng = np.random.default_rng(0)
pp = rng.standard_normal(size=(4, 100, 10))
qs = np.quantile(pp, np.linspace(0, 1, 10))
obs = qs[np.arange(10), np.arange(10)]
obs[0] *= .9
obs[-1] *= 1.1
idata = az.from_dict(
    posterior_predictive={"y": pp},
    observed_data={"y": obs},
    log_likelihood={"y": -pp**2, "decoy": np.zeros_like(pp)}
)
loo_pit_data = loo_pit(idata, y="y")
assert np.all((loo_pit_data >= 0) & (loo_pit_data <= 1))

[utkarsh-maheshwari]

idata = az.from_dict(
posterior_predictive={"y": pp},
observed_data={"y": obs},
log_likelihood={"y": -pp**2, "decoy": np.zeros_like(pp)}
)

@OriolAbril I guess loo_pit need the posterior as well?

arviz/arviz/stats/stats.py

Line 1588 in 4d9caca

posterior = convert_to_dataset(idata, group="posterior")

[OriolAbril]

Oh yes, it should have some random samples.

[utkarsh-maheshwari]

@canyon289 I can take this. This PR solves other issues as well.!

[utkarsh-maheshwari]

Thanks for helping with the dataset 😄 . @OriolAbril

[utkarsh-maheshwari]

I did'nt mentioned the changes in changelog. Thought of doing it after seeing all green

[OriolAbril]

Feel free to add them to the changelog in one of your other open PRs