aai-institute / pyDVL

pyDVL is a library of stable implementations of algorithms for data valuation and influence function computation
https://pydvl.org
GNU Lesser General Public License v3.0
109 stars 8 forks source link

Refactor Classwise Shapley #616

Closed AnesBenmerzoug closed 3 months ago

AnesBenmerzoug commented 3 months ago

Description

This PR refactors the Classwise Shapley valuation to be more aligned with the new Valuation classes.

Changes

Checklist

AnesBenmerzoug commented 3 months ago

@schroedk I resolved as many issues as possible but there is still one test that fails for some unknown reason and I am giving up on it. The test in question relies on non-default behaviour, i.e., it does not normalize the values after the valuation. The original paper's implementation normalizes the values and for that case the test passes.

schroedk commented 3 months ago

@schroedk I resolved as many issues as possible but there is still one test that fails for some unknown reason and I am giving up on it. The test in question relies on non-default behaviour, i.e., it does not normalize the values after the valuation. The original paper's implementation normalizes the values and for that case the test passes.

@AnesBenmerzoug the test is also failing on #558. Let's go for merging this branch and then iterate on #558. What is missing before we can merge this one?

AnesBenmerzoug commented 3 months ago

@schroedk the test I was talking about is specific to classwise shapley. I have no idea about that other test that is failing.

I think this PR can now be merged.