For our package http://github.com/metric-learn/metric-learn, the function CalibratedClassifierCV is very convenient for Weakly Supervised Learners, as it can make PairsClassifier estimators return a probability for a pair of points to be labeled as similar or dissimilar, when those return a decision function.
However, we currently cannot use it because our inputs can be 3D (example: pairs=[[[2.3, 5.4], [4.4, 5.6]], [[7.5, 1.2], [4.4, 5.6]]]), and CalibratedClassifierCV uses check_array with default parameters that does not allow 3D inputs.
However, other meta-estimators like GridSearchCV do not call check_array, so we can use them easily in metric-learn.
Is the check_array in CalibratedClassifierCV really useful or could we do without it ? If we could, I'd be happy to provide a PR to do so
For our package http://github.com/metric-learn/metric-learn, the function
CalibratedClassifierCVis very convenient for Weakly Supervised Learners, as it can make PairsClassifier estimators return a probability for a pair of points to be labeled as similar or dissimilar, when those return a decision function.However, we currently cannot use it because our inputs can be 3D (example:
pairs=[[[2.3, 5.4], [4.4, 5.6]], [[7.5, 1.2], [4.4, 5.6]]]), andCalibratedClassifierCVusescheck_arraywith default parameters that does not allow 3D inputs.However, other meta-estimators like
GridSearchCVdo not callcheck_array, so we can use them easily in metric-learn.Is the
check_arrayinCalibratedClassifierCVreally useful or could we do without it ? If we could, I'd be happy to provide a PR to do so