handle I;16 mode in pil_to_tensor

docs/source/transforms.rst

@@ -99,6 +99,15 @@ have values in ``[0, MAX_DTYPE]`` where ``MAX_DTYPE`` is the largest value
 that can be represented in that dtype. Typically, images of dtype
 ``torch.uint8`` are expected to have values in ``[0, 255]``.
 
+.. note::
+
+    ``torch.uint16``, ``torch.uint32``, and ``torch.uint64`` dtypes are not
+    officially supported by the torchvision transforms. While some operations
+    may work, most transforms expect ``torch.uint8`` or ``torch.float32``
+    inputs. If you're working with uint16 images (e.g. from 16-bit medical or
+    scientific imaging), consider converting to ``float32`` first using
+    :class:`~torchvision.transforms.v2.ToDtype`.
+
 Use :class:`~torchvision.transforms.v2.ToDtype` to convert both the dtype and
 range of the inputs.
 

test/test_transforms.py

@@ -614,8 +614,8 @@ def _get_1_channel_tensor_various_types():
         expected_output = img_data_byte.float().div(255.0).numpy()
         yield img_data_byte, expected_output, "L"
 
-        img_data_short = torch.ShortTensor(1, 4, 4).random_()
-        expected_output = img_data_short.numpy()
+        img_data_short = torch.ShortTensor(1, 4, 4).random_(0, 32767)
+        expected_output = img_data_short.numpy().astype(np.uint16)
         yield img_data_short, expected_output, "I;16" if sys.byteorder == "little" else "I;16B"
 
         img_data_int = torch.IntTensor(1, 4, 4).random_()
@@ -631,8 +631,8 @@ def _get_2d_tensor_various_types():
         expected_output = img_data_byte.float().div(255.0).numpy()
         yield img_data_byte, expected_output, "L"
 
-        img_data_short = torch.ShortTensor(4, 4).random_()
-        expected_output = img_data_short.numpy()
+        img_data_short = torch.ShortTensor(4, 4).random_(0, 32767)
+        expected_output = img_data_short.numpy().astype(np.uint16)
         yield img_data_short, expected_output, "I;16" if sys.byteorder == "little" else "I;16B"
 
         img_data_int = torch.IntTensor(4, 4).random_()

test/test_transforms_v2.py

@@ -6879,6 +6879,18 @@ def test_functional_error(self):
             F.pil_to_tensor(object())
 
 
+@pytest.mark.parametrize("f", [F.to_tensor, F.pil_to_tensor])
+def test_I16_to_tensor(f):
+    # See https://github.com/pytorch/vision/issues/8188
+    I16_pil_img = PIL.Image.fromarray(np.random.randint(0, 2**16, (10, 10), dtype=np.uint16))
+    assert I16_pil_img.mode == "I;16"
+
+    cm = pytest.warns(UserWarning, match="deprecated") if f is F.to_tensor else contextlib.nullcontext()
+    with cm:
+        out = f(I16_pil_img)
+    assert out.dtype == torch.uint16
+
+
 @needs_cvcuda
 class TestToCVCUDATensor:
     @pytest.mark.parametrize("image_type", (torch.Tensor, tv_tensors.Image))

torchvision/transforms/functional.py

@@ -164,7 +164,7 @@ def to_tensor(pic: Union[PILImage, np.ndarray]) -> Tensor:
         return torch.from_numpy(nppic).to(dtype=default_float_dtype)
 
     # handle PIL Image
-    mode_to_nptype = {"I": np.int32, "I;16" if sys.byteorder == "little" else "I;16B": np.int16, "F": np.float32}
+    mode_to_nptype = {"I": np.int32, "I;16" if sys.byteorder == "little" else "I;16B": np.uint16, "F": np.float32}
     img = torch.from_numpy(np.array(pic, mode_to_nptype.get(pic.mode, np.uint8), copy=True))
 
     if pic.mode == "1":