mitigation methods with image/multi-dimension data

[nina-weng]

I am trying to use the fairlearn bias mitigation methods with multi-dimension data like images. To my understanding, the postprocessing methods like ThresholdOptimizer, should not be constraint by the type of classifiers. However, when I try to implement the ThresholdOptimizer with input X (shape:[BS, C, W, H]), it keeps giving me error like:

ValueError: Found array with dim 4. None expected <= 2.

Which, if I understand correctly, suggests that X should only have a dim <=2.

I try to wrapped the clf model based on sklearn estimator, but it does not help as well.

I was wondering, how could I solve this issue?

---

version
scikit-learn 1.3.0

---

The error is coming from this line
if not allow_nd and array.ndim >= 3: raise ValueError( "Found array with dim %d. %s expected <= 2." % (array.ndim, estimator_name) )
in sklearn/utils/validation.py:953.
I belive that in fairlearn/utils/_input_validation.py:89:
X, y = check_X_y(X, y, dtype=None, force_all_finite=False), we should add allow_nd = True.

[alliesaizan]

Hi @nina-weng! More information would help us troubleshoot. Can you please post a screenshot of the code you are working with here, including the parts with the unmitigated estimator and the issue with Threshold Optimizer?

[adrinjalali]

I think I agree with @nina-weng here.

@nina-weng could you please open a PR to fix the issue?