You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
fkiraly
changed the title
[BUG] classifiers failing on multiclass scenario
[BUG] classifiers failing on multiclass scenario due to _get_train_probsMay 2, 2024
Update: this is due to the _get_train_probs legacy method which has an undefined contract but which seems to assume a narrow range of data types (e.g., 3D numpy).
…data type (#6377)
This PR fixes#6376 by ensuring the `_get_train_probs` method - with
undocumented contract - accepts `X_train` of any permissible panel input
type.
The fix is adding an input converter to numpy3D, which all currently
implemented instances seem to assume.
This approach should fix the failing `test_classifier_output` test for
the new scenario, without removing functionality (even if private), or
degrading efficiency in a case where the input is numpy already.
Depends on #6374 for testing.
The following classifiers fail on a scenario with three classes, see #6374 for the scenario:
Arsenal
BOSSEnsemble
ContractableBOSS
DrCIF
FreshPRINCE
ShapeletTransformClassifier
TemporalDictionaryEnsemble
TSCGridSearchCV
- reason is not the classifier, but folds being too small in the scenarioWeightedEnsembleClassifier
A quick solution could be to introduce a
capability:multiclass
tag, but perhaps the bug is easy to fix.The text was updated successfully, but these errors were encountered: