🔨 Finish reworking CUB

torch_uncertainty/datasets/classification/cub.py

@@ -21,7 +21,7 @@ def __init__(
         train: bool = True,
         transform: Callable | None = None,
         target_transform: Callable | None = None,
-        load_attributes: bool = False,
+        return_attributes: bool = False,
         download: bool = False,
     ):
         """The Caltech-UCSD Birds-200-2011 dataset.
@@ -34,8 +34,8 @@ def __init__(
                 returns a transformed version. E.g, transforms.RandomCrop. Defaults to None.
             target_transform (callable, optional): A function/transform that takes in the target
                 and transforms it. Defaults to None.
-            load_attributes (bool, optional): If True, loads the attributes of the dataset and
-                returns them instead of the images. Defaults to False.
+            return_attributes (bool, optional): If True, returns the attributes instead of the images.
+                Defaults to False.
             download (bool, optional): If True, downloads the dataset from the internet and puts it
                 in root directory. If dataset is already downloaded, it is not downloaded again.
                 Defaults to False.
@@ -57,16 +57,26 @@ def __init__(
         super().__init__(Path(root) / "CUB_200_2011" / "images", transform, target_transform)
 
         training_idx = self._load_train_idx()
-        self.attributes, self.certainties = self._load_attributes()
-        if load_attributes:
-            self.samples = zip(self.attributes, [sam[1] for sam in self.samples], strict=False)
-            self.attribute_names = self._load_attribute_names()
-            self.loader = torch.nn.Identity()
+        self.attributes, self.uncertainties = self._load_attributes()
+        self.attribute_names = self._load_attribute_names()
+        self.classnames = self._load_classnames()
 
         self.samples = [sample for i, sample in enumerate(self.samples) if training_idx[i] == train]
         self._labels = [label for i, label in enumerate(self.targets) if training_idx[i] == train]
+        self.attributes = rearrange(
+            torch.masked_select(self.attributes, training_idx.unsqueeze(-1) == train),
+            "(n c) -> n c",
+            c=312,
+        )
+        self.uncertainties = rearrange(
+            torch.masked_select(self.uncertainties, training_idx.unsqueeze(-1) == train),
+            "(n c) -> n c",
+            c=312,
+        )
 
-        self.classnames = self._load_classnames()
+        if return_attributes:
+            self.samples = zip(self.attributes, [sam[1] for sam in self.samples], strict=False)
+            self.loader = torch.nn.Identity()
 
     def _load_classnames(self) -> list[str]:
         """Load the classnames of the dataset.
@@ -92,16 +102,18 @@ def _load_attributes(self) -> tuple[Tensor, Tensor]:
         """Load the attributes associated to each image.
 
         Returns:
-            tuple[Tensor, Tensor]: The presence of the 312 attributes along with their certainty.
+            tuple[Tensor, Tensor]: The presence of the 312 attributes along with their uncertainty.
+                The uncertainty is 0 for certain samples and 1 for non-visible attributes.
         """
-        attributes, certainty = [], []
+        attributes, uncertainty = [], []
         with (self.folder_root / "CUB_200_2011" / "attributes" / "image_attribute_labels.txt").open(
             "r"
         ) as f:
-            attributes = [int(line.split(" ")[2]) for line in f]
-            certainty = [(int(line.split(" ")[3]) - 1) / 3 for line in f]
+            for line in f:
+                attributes.append(int(line.split(" ")[2]))
+                uncertainty.append(1 - (int(line.split(" ")[3]) - 1) / 3)
         return rearrange(torch.as_tensor(attributes), "(n c) -> n c", c=312), rearrange(
-            torch.as_tensor(certainty), "(n c) -> n c", c=312
+            torch.as_tensor(uncertainty), "(n c) -> n c", c=312
         )
 
     def _load_attribute_names(self) -> list[str]: