Add torchvision.transforms v1/v2 independent ToTensor() implementation (

#1718)

* add ToTensor() implementation compatible with v2

* make version independent ToTensor available

* add test for version indep. ToTensor() implementation

lightly/transforms/__init__.py

@@ -46,7 +46,10 @@
     TiCoView1Transform,
     TiCoView2Transform,
 )
-from lightly.transforms.torchvision_transforms import torchvision_transforms
+from lightly.transforms.torchvision_v2_compatibility import (
+    ToTensor,
+    torchvision_transforms,
+)
 from lightly.transforms.vicreg_transform import VICRegTransform, VICRegViewTransform
 from lightly.transforms.vicregl_transform import VICRegLTransform, VICRegLViewTransform
 from lightly.transforms.wmse_transform import WMSETransform

lightly/transforms/torchvision_v2_compatibility.py

@@ -0,0 +1,49 @@
+#
+# Copyright (c) Lightly AG and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+#
+from typing import Union
+
+import torch
+from PIL.Image import Image
+from torch import Tensor
+from torchvision.transforms import ToTensor as ToTensorV1
+
+try:
+    from torchvision.transforms import v2 as torchvision_transforms
+
+    _TRANSFORMS_V2 = True
+
+except ImportError:
+    from torchvision import transforms as torchvision_transforms
+
+    _TRANSFORMS_V2 = False
+
+
+def ToTensor() -> Union[torchvision_transforms.Compose, ToTensorV1]:
+    """Convert a PIL Image to a tensor with value normalization, similar to [0].
+
+    This implementation is required since `torchvision.transforms.v2.ToTensor` is
+    deprecated and will be removed in the future (see [1]).
+
+    Input to this transform:
+        PIL Image (H x W x C) of uint8 type in range [0,255]
+
+    Output of this transform:
+        torch.Tensor (C x H x W) of type torch.float32 in range [0.0, 1.0]
+
+    - [0] https://pytorch.org/vision/main/generated/torchvision.transforms.ToTensor.html
+    - [1] https://pytorch.org/vision/0.20/generated/torchvision.transforms.v2.ToTensor.html?highlight=totensor#torchvision.transforms.v2.ToTensor
+    """
+    T = torchvision_transforms
+    if _TRANSFORMS_V2 and hasattr(T, "ToImage") and hasattr(T, "ToDtype"):
+        # v2.transforms.ToTensor is deprecated and will be removed in the future.
+        # This is the new recommended way to convert a PIL Image to a tensor since
+        # torchvision v0.16.
+        # See also https://github.com/pytorch/vision/blame/33e47d88265b2d57c2644aad1425be4fccd64605/torchvision/transforms/v2/_deprecated.py#L19
+        return T.Compose([T.ToImage(), T.ToDtype(dtype=torch.float32, scale=True)])
+    else:
+        return T.ToTensor()

tests/transforms/test_torchvision_v2compatibility.py

@@ -0,0 +1,16 @@
+import numpy as np
+import torch
+from PIL import Image
+from torch import Tensor
+
+from lightly.transforms import ToTensor
+
+
+def test_ToTensor() -> None:
+    img_np = np.random.randint(0, 255, (20, 30, 3), dtype=np.uint8)
+    img_pil = Image.fromarray(img_np)
+    img_tens = ToTensor()(img_pil)
+    assert isinstance(img_tens, Tensor)
+    assert img_tens.shape == (3, 20, 30)
+    assert img_tens.dtype == torch.float32
+    assert img_tens.max() <= 1.0 and img_tens.min() >= 0.0