add warning for CelebA bbox data (#9339)

ribbon-otter · ribbon-otter · zy1git · web-flow · commit 6940e19087ce · 2026-02-09T09:47:14.000-08:00
Co-authored-by: ribbon-otter &lt;ribbon-otter@github.invalid&gt;
Co-authored-by: zy1git &lt;zycoding1@gmail.com&gt;
Co-authored-by: Nicolas Hug &lt;contact@nicolas-hug.com&gt;
Co-authored-by: Zhitao Yu &lt;zhitao@fb.com&gt;
Co-authored-by: Nicolas Hug &lt;nh.nicolas.hug@gmail.com&gt;
diff --git a/torchvision/datasets/celeba.py b/torchvision/datasets/celeba.py
@@ -26,7 +26,18 @@ class CelebA(VisionDataset):
 
                 - ``attr`` (Tensor shape=(40,) dtype=int): binary (0, 1) labels for attributes
                 - ``identity`` (int): label for each person (data points with the same identity are the same person)
-                - ``bbox`` (Tensor shape=(4,) dtype=int): bounding box (x, y, width, height)
+                - ``bbox`` (Tensor shape=(4,) dtype=int): bounding box (x, y, width, height).
+
+                  .. warning::
+
+                      These bounding box coordinates correspond to the original uncropped
+                      CelebA images, not the cropped and aligned images returned by this
+                      dataset. As a result, the coordinates will not match and may fall
+                      outside the image boundaries.
+
+                      See `Issue #9008 <https://github.com/pytorch/vision/issues/9008>`_ for
+                      details and potential workarounds.
+
                 - ``landmarks`` (Tensor shape=(10,) dtype=int): landmark points (lefteye_x, lefteye_y, righteye_x,
                   righteye_y, nose_x, nose_y, leftmouth_x, leftmouth_y, rightmouth_x, rightmouth_y)
 
