Torchvision API to tensor/PIL image conversion operators

dali/python/nvidia/dali/experimental/torchvision/__init__.py

@@ -20,6 +20,7 @@
 from .v2.normalize import Normalize
 from .v2.pad import Pad
 from .v2.resize import Resize
+from .v2.totensor import ToPureTensor, PILToTensor, ToPILImage
 
 __all__ = [
     "CenterCrop",
@@ -29,7 +30,10 @@
     "Grayscale",
     "Normalize",
     "Pad",
+    "PILToTensor",
     "RandomHorizontalFlip",
     "RandomVerticalFlip",
     "Resize",
+    "ToPILImage",
+    "ToPureTensor",
 ]

dali/python/nvidia/dali/experimental/torchvision/v2/compose.py

@@ -22,6 +22,8 @@
 from nvidia.dali.backend import TensorListCPU, TensorListGPU
 
 from .operator import _ValidateTensorOrImage
+from .totensor import ToPureTensor, PILToTensor, ToPILImage
+from .functional.totensor import to_pil_image
 
 import numpy as np
 import multiprocessing
@@ -119,6 +121,17 @@ def _cuda_run(self, data_input):
     def _cpu_run(self, data_input):
         return self.pipe.run(input_data=data_input)
 
+    @staticmethod
+    def _get_output_format(op_list: List[Callable[..., Sequence[_DataNode] | _DataNode]]) -> str:
+        output_type = "default"
+        for op in op_list:
+            if isinstance(op, (ToPureTensor, PILToTensor)):
+                output_type = "tensor"
+            elif isinstance(op, ToPILImage):
+                output_type = "pil"
+
+        return output_type
+
     def __init__(
         self,
         op_list: List[Callable[..., Sequence[_DataNode] | _DataNode]],
@@ -127,16 +140,7 @@ def __init__(
         num_threads: int = DEFAULT_NUM_THREADS,
         **dali_pipeline_kwargs,
     ):
-        # TODO:
-        # convert_to_tensor is currently not supported and requires an user's effort
-        # to convert to tensor
-        # ToTensor is deprecated and according to:
-        # https://docs.pytorch.org/vision/stable/_modules/torchvision/transforms/v2/_deprecated.html#ToTensor
-        # should be replaced with:
-        # v2.Compose([v2.ToImage(), v2.ToDtype(torch.float32, scale=True)])
-        #
-        # self.convert_to_tensor = True if isinstance(op_list[-1], ToTensor) else False
-        self.convert_to_tensor = False
+        self.output_type = PipelineWithLayout._get_output_format(op_list)
         self.device = op_list[0].device if len(op_list) > 0 else "cpu"
         self.torch_device_type = "cuda" if self.device == "gpu" else "cpu"
 
@@ -167,14 +171,29 @@ def run(self, data_input):
         if output is None:
             return output
 
-        output = to_torch_tensor(output)
-        # ToTensor
-        if self.convert_to_tensor:
-            if output.shape[-4] > 1:
-                raise NotImplementedError("ToTensor does not currently work for batches")
+        return to_torch_tensor(output)
 
+    def output_to_tensor(self, output: torch.Tensor) -> torch.Tensor:
+        """Return the pipeline output as a CHW or NCHW ``torch.Tensor``.
+
+        For HWC pipelines the axes are permuted; for CHW pipelines the tensor is returned
+        as-is (batch dimension handling is left to the caller).
+        """
+        if self.get_layout() == "HWC":
+            # output shape: (N, H, W, C)
+            if output.shape[0] == 1:
+                return output.squeeze(0).permute(2, 0, 1)  # → (C, H, W)
+            return output.permute(0, 3, 1, 2)  # → (N, C, H, W)
+        # CHW — already the right layout; batch handling belongs to the subclass
         return output
 
+    def output_to_pil(self, output: torch.Tensor) -> Image.Image:
+        """Convert a single-sample pipeline output tensor to a ``PIL.Image``.
+
+        For CHW pipelines the tensor is permuted to HWC before conversion.
+        """
+        return to_pil_image(output)
+
     @abstractmethod
     def get_layout(self) -> str: ...
 
@@ -185,7 +204,7 @@ def get_channel_reverse_idx(self) -> int: ...
     def verify_layout(self, data) -> None: ...
 
     def is_conversion_to_tensor(self) -> bool:
-        return self.convert_to_tensor
+        return self.output_type == "tensor"
 
 
 class PipelineHWC(PipelineWithLayout):
@@ -221,30 +240,6 @@ def __init__(
             **dali_pipeline_kwargs,
         )
 
-    def _convert_tensor_to_image(self, in_tensor: torch.Tensor):
-
-        channels = self.get_channel_reverse_idx()
-
-        # TODO: consider when to convert to PIL.Image - e.g. if it make sense for channels < 3
-        # There is no certain method to determine if the tensor is HW, HWC, or NHWC.
-        # The method below checks if tensor's shape is HW or ...HWC with a single channel
-        if len(in_tensor.shape) == 2 or (
-            len(in_tensor.shape) >= 3 and in_tensor.shape[channels] == 1
-        ):
-            mode = "L"
-            if len(in_tensor.shape) != 2:
-                in_tensor = in_tensor.squeeze(-1)
-        elif in_tensor.shape[channels] == 3:
-            mode = "RGB"
-        elif in_tensor.shape[channels] == 4:
-            mode = "RGBA"
-        else:
-            raise ValueError(
-                f"Unsupported number of channels: {in_tensor.shape[channels]}. Should be 1, 3 or 4."
-            )
-        # We need to convert tensor to CPU, PIL does not support CUDA tensors
-        return Image.fromarray(in_tensor.cpu().numpy(), mode=mode)
-
     def run(self, data_input):
         if isinstance(data_input, Image.Image):
             _input = torch.as_tensor(np.array(data_input, copy=True)).unsqueeze(0)
@@ -254,24 +249,20 @@ def run(self, data_input):
             raise ValueError("HWC layout is currently supported for PIL Images only.\
                 Please check if samples have the same format.")
 
-        output = super().run(_input)
+        output = super().run(_input)  # (N, H, W, C)
 
-        if self.is_conversion_to_tensor():
-            return output
+        if self.output_type == "tensor":
+            return self.output_to_tensor(output)
 
-        if isinstance(output, tuple):
-            output = self._convert_tensor_to_image(output[0])
-        else:
-            # batches
-            if output.shape[0] > 1:
-                output_list = []
-                for i in range(output.shape[0]):
-                    output_list.append(self._convert_tensor_to_image(output[i]))
-                output = output_list
-            else:
-                output = self._convert_tensor_to_image(output[0])
+        # default: PIL output
+        if output.shape[0] > 1:
+            return [self.output_to_pil(output[i]) for i in range(output.shape[0])]
+        return self.output_to_pil(output[0])
 
-        return output
+    def output_to_pil(self, output: torch.Tensor) -> Image.Image:
+        """Convert a single-sample HWC (H, W, C) tensor to a ``PIL.Image``."""
+        # Pipeline stores data as HWC; to_pil_image expects CHW, so permute first.
+        return to_pil_image(output.permute(2, 0, 1))
 
     def get_layout(self) -> str:
         return "HWC"
@@ -331,6 +322,14 @@ def run(self, data_input):
         if data_input.ndim == 3:
             # Remove the batch dimension we added above
             output = output.squeeze(0)
+
+        if self.output_type == "pil":
+            if output.ndim == 4:
+                if output.shape[0] > 1:
+                    return [self.output_to_pil(output[i]) for i in range(output.shape[0])]
+                return self.output_to_pil(output[0])
+            return self.output_to_pil(output)
+
         return output
 
     def get_layout(self) -> str:

dali/python/nvidia/dali/experimental/torchvision/v2/functional/__init__.py

@@ -19,15 +19,19 @@
 from .normalize import normalize
 from .pad import pad
 from .resize import resize
+from .totensor import pil_to_tensor, to_tensor, to_pil_image
 
 __all__ = [
     "center_crop",
     "gaussian_blur",
     "horizontal_flip",
     "normalize",
     "pad",
+    "pil_to_tensor",
     "resize",
     "rgb_to_grayscale",
     "to_grayscale",
+    "to_pil_image",
+    "to_tensor",
     "vertical_flip",
 ]

dali/python/nvidia/dali/experimental/torchvision/v2/functional/totensor.py

@@ -0,0 +1,116 @@
+# Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from PIL import Image
+import torch
+
+
+def pil_to_tensor(inpt: Image.Image | np.ndarray) -> torch.Tensor:
+    """
+    Convert a ``PIL.Image`` to a uint8 CHW ``torch.Tensor``.
+
+    Values are in [0, 255] and the dtype is ``torch.uint8``.  No scaling is applied.
+    Mirrors ``torchvision.transforms.v2.functional.pil_to_tensor``.
+
+    Parameters
+    ----------
+    inpt : PIL.Image
+        Input image.  Modes ``L``, ``RGB``, and ``RGBA`` are supported.
+
+    Returns
+    -------
+    torch.Tensor
+        CHW tensor of dtype matching PIL mode (uint8 for L/RGB/RGBA, int32 for I, float32 for F).
+    """
+    if not isinstance(inpt, (Image.Image, np.ndarray)):
+        raise TypeError(f"Expected PIL.Image or numpy array, got {type(inpt)}")
+
+    if isinstance(inpt, Image.Image):
+        arr = np.array(inpt, copy=True)  # (H, W) for L, (H, W, C) for RGB/RGBA
+    else:
+        # Note: numpy array is used directly without copying, this matches Torchvision
+        arr = inpt
+
+    if arr.ndim == 2:
+        arr = np.expand_dims(arr, axis=-1)  # (H, W) → (H, W, 1)
+
+    return torch.from_numpy(arr).permute(2, 0, 1)  # (H, W, C) → (C, H, W)
+
+
+def to_tensor(inpt: Image.Image | np.ndarray) -> torch.Tensor:
+    """
+    Convert a ``PIL.Image`` to a float32 CHW ``torch.Tensor`` with values in [0, 1].
+
+    Mirrors ``torchvision.transforms.v2.functional.to_tensor`` (deprecated in TV v2,
+    but kept here for compatibility).
+
+    Parameters
+    ----------
+    inpt : PIL.Image
+        Input image.  Modes ``L``, ``RGB``, and ``RGBA`` are supported.
+
+    Returns
+    -------
+    torch.Tensor
+        CHW tensor of dtype ``torch.float32`` with values in [0.0, 1.0].
+    """
+    return pil_to_tensor(inpt).float() / 255.0
+
+
+def to_pil_image(inpt: torch.Tensor, mode: str | None = None) -> Image.Image:
+    """
+    Convert a CHW ``torch.Tensor`` to a ``PIL.Image``.
+
+    Mirrors ``torchvision.transforms.v2.functional.to_pil_image``.
+
+    Parameters
+    ----------
+    inpt : torch.Tensor
+        CHW tensor.  Supported channel counts: 1 (``L``), 3 (``RGB``), 4 (``RGBA``).
+    mode : str or None, optional
+        PIL image mode.  If ``None`` the mode is inferred from the channel count.
+
+    Returns
+    -------
+    PIL.Image
+    """
+    if not isinstance(inpt, torch.Tensor):
+        raise TypeError(f"Expected torch.Tensor, got {type(inpt)}")
+    if inpt.ndim != 3:
+        raise ValueError(f"Expected 3-D CHW tensor, got shape {tuple(inpt.shape)}")
+
+    hwc = inpt.permute(1, 2, 0).cpu()  # (C, H, W) → (H, W, C)
+    channels = hwc.shape[-1]
+
+    if mode is None:
+        if channels == 1:
+            mode = "L"
+        elif channels == 3:
+            mode = "RGB"
+        elif channels == 4:
+            mode = "RGBA"
+        else:
+            raise ValueError(
+                f"Cannot infer PIL mode from {channels} channels. " "Pass mode explicitly."
+            )
+
+    arr = hwc.numpy()
+    if np.issubdtype(arr.dtype, np.floating) and mode != "F":
+        arr = (arr * 255).astype(np.uint8)
+
+    if mode == "L":
+        arr = arr.squeeze(-1)
+
+    return Image.fromarray(arr, mode=mode)

dali/python/nvidia/dali/experimental/torchvision/v2/operator.py

@@ -270,7 +270,7 @@ def transform_input(inpt, device: Literal["cpu", "gpu"] = "cpu") -> ndd.Tensor |
             if inpt.ndim == 3:
                 _input = ndd.Tensor(inpt, layout="CHW")
             elif inpt.ndim > 3:
-                # Creating baches of NCHW
+                # Creating batches of NCHW
                 _input = ndd.as_batch(inpt, layout="CHW")
             else:
                 raise TypeError(f"Tensor has < 3 dimensions: {inpt.ndim}, shape: {inpt.shape}")