Merge pull request #2022 from Trusted-AI/development_issue_2014

Add conversion to float for comparison on GPU/CUDA

Author: beat-buesser

art/estimators/classification/pytorch.py

@@ -272,7 +272,7 @@ def reduce_labels(self, y: Union[np.ndarray, "torch.Tensor"]) -> Union[np.ndarra
         # Check if the loss function supports probability labels and probability labels are provided
         if self._probability_labels and len(y.shape) == 2:
             if isinstance(y, torch.Tensor):
-                is_one_hot = torch.equal(y.floor(), y.ceil())
+                is_one_hot = torch.equal(y.float().floor(), y.float().ceil())
             else:
                 is_one_hot = np.array_equal(np.floor(y), np.ceil(y))
             if not is_one_hot:  # probability labels