[Bugfix] Remove incorrect torchvision requirement from PIL backend image processors

src/transformers/image_processing_backends.py

@@ -529,7 +529,7 @@ def resize(
         self,
         image: np.ndarray,
         size: SizeDict,
-        resample: Union["PILImageResampling", "tvF.InterpolationMode", int] | None = None,
+        resample: "PILImageResampling | None" = None,
         reducing_gap: int | None = None,
         **kwargs,
     ) -> np.ndarray:
@@ -628,7 +628,7 @@ def _preprocess(
         images: list[np.ndarray],
         do_resize: bool,
         size: SizeDict,
-        resample: Union["PILImageResampling", "tvF.InterpolationMode", int] | None,
+        resample: "PILImageResampling | None",
         do_center_crop: bool,
         crop_size: SizeDict,
         do_rescale: bool,

src/transformers/models/aria/image_processing_aria.py

@@ -18,17 +18,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import torch
+from torchvision.transforms.v2 import functional as tvF
 
 from ...image_processing_backends import TorchvisionBackend
 from ...image_processing_utils import BatchFeature, get_patch_output_size, select_best_resolution
 from ...image_transforms import divide_to_patches
 from ...image_utils import ChannelDimension, PILImageResampling, SizeDict, get_image_size
 from ...processing_utils import ImagesKwargs, Unpack
-from ...utils import TensorType, auto_docstring, is_torchvision_available
-
-
-if is_torchvision_available():
-    from torchvision.transforms.v2 import functional as tvF
+from ...utils import TensorType, auto_docstring
 
 
 class AriaImageProcessorKwargs(ImagesKwargs, total=False):

src/transformers/models/aria/image_processing_pil_aria.py

@@ -24,13 +24,29 @@
     SizeDict,
     get_image_size,
 )
-from ...processing_utils import Unpack
+from ...processing_utils import ImagesKwargs, Unpack
 from ...utils import TensorType, auto_docstring
-from ...utils.import_utils import requires
-from .image_processing_aria import AriaImageProcessorKwargs
 
 
-@requires(backends=("vision", "torch", "torchvision"))
+# Adapted from transformers.models.aria.image_processing_aria.AriaImageProcessorKwargs
+class AriaImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    max_image_size (`int`, *optional*, defaults to `self.max_image_size`):
+        Maximum image size. Must be either 490 or 980.
+    min_image_size (`int`, *optional*, defaults to `self.min_image_size`):
+        Minimum image size. Images smaller than this in any dimension will be scaled up.
+    split_resolutions (`list[list[int]]`, *optional*, defaults to `self.split_resolutions`):
+        A list of possible resolutions as (height, width) pairs for splitting high-resolution images into patches.
+    split_image (`bool`, *optional*, defaults to `self.split_image`):
+        Whether to split the image into patches using the best matching resolution from `split_resolutions`.
+    """
+
+    max_image_size: int
+    min_image_size: int
+    split_resolutions: list[list[int]]
+    split_image: bool
+
+
 @auto_docstring
 class AriaImageProcessorPil(PilBackend):
     model_input_names = ["pixel_values", "pixel_mask", "num_crops"]
@@ -65,7 +81,7 @@ def _resize_for_patching(
         self,
         image: np.ndarray,
         target_resolution: tuple,
-        resample: "PILImageResampling | int | None",
+        resample: "PILImageResampling | None",
     ) -> np.ndarray:
         """Resize an image to a target resolution while maintaining aspect ratio."""
         new_height, new_width = get_patch_output_size(
@@ -90,7 +106,7 @@ def get_image_patches(
         image: np.ndarray,
         grid_pinpoints: list[list[int]],
         patch_size: int,
-        resample: "PILImageResampling | int | None",
+        resample: "PILImageResampling | None",
     ) -> list[np.ndarray]:
         """
         Process an image with variable resolutions by dividing it into patches.
@@ -131,7 +147,7 @@ def _preprocess(
         min_image_size: int = 336,
         split_resolutions: list[list[int]] | None = None,
         split_image: bool = False,
-        resample: "PILImageResampling | int | None" = None,
+        resample: "PILImageResampling | None" = None,
         **kwargs,
     ) -> BatchFeature:
         if max_image_size not in [490, 980]:

src/transformers/models/aria/modeling_aria.py

@@ -41,12 +41,7 @@
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import (
-    TransformersKwargs,
-    auto_docstring,
-    can_return_tuple,
-    torch_compilable_check,
-)
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, torch_compilable_check
 from ...utils.generic import maybe_autocast, merge_with_config_defaults
 from ...utils.output_capturing import capture_outputs
 from ..auto import AutoModel

src/transformers/models/aria/modular_aria.py

@@ -14,6 +14,7 @@
 import torch
 from huggingface_hub.dataclasses import strict
 from torch import nn
+from torchvision.transforms.v2 import functional as tvF
 
 from ... import initialization as init
 from ...activations import ACT2FN
@@ -39,8 +40,6 @@
     TransformersKwargs,
     auto_docstring,
     can_return_tuple,
-    is_torch_available,
-    is_torchvision_available,
     logging,
 )
 from ..auto import CONFIG_MAPPING, AutoConfig, AutoTokenizer
@@ -323,13 +322,6 @@ def forward(self, key_value_states: torch.Tensor, attn_mask: torch.Tensor | None
         return out
 
 
-if is_torch_available():
-    import torch
-
-if is_torchvision_available():
-    from torchvision.transforms.v2 import functional as tvF
-
-
 class AriaImageProcessorKwargs(ImagesKwargs, total=False):
     r"""
     max_image_size (`int`, *optional*, defaults to `self.max_image_size`):

src/transformers/models/beit/image_processing_beit.py

@@ -15,6 +15,10 @@
 
 from typing import Union
 
+import torch
+import torch.nn.functional as F
+from torchvision.transforms.v2 import functional as tvF
+
 from ...image_processing_backends import TorchvisionBackend
 from ...image_processing_utils import BatchFeature
 from ...image_transforms import group_images_by_shape, reorder_images
@@ -27,15 +31,7 @@
     SizeDict,
 )
 from ...processing_utils import ImagesKwargs, Unpack
-from ...utils import TensorType, auto_docstring, is_torch_available, is_torchvision_available
-
-
-if is_torch_available():
-    import torch
-    import torch.nn.functional as F
-
-if is_torchvision_available():
-    from torchvision.transforms.v2 import functional as tvF
+from ...utils import TensorType, auto_docstring, is_torch_available
 
 
 class BeitImageProcessorKwargs(ImagesKwargs, total=False):

src/transformers/models/beit/image_processing_pil_beit.py

@@ -14,9 +14,6 @@
 """Image processor class for BEiT."""
 
 import numpy as np
-import torch
-import torch.nn.functional as F
-from torchvision.transforms.v2 import functional as tvF
 
 from ...image_processing_backends import PilBackend
 from ...image_processing_utils import BatchFeature
@@ -28,14 +25,24 @@
     PILImageResampling,
     SizeDict,
 )
-from ...processing_utils import Unpack
-from ...utils import TensorType, auto_docstring, is_torch_available
+from ...processing_utils import ImagesKwargs, Unpack
+from ...utils import TensorType, auto_docstring
 from ...utils.import_utils import requires
-from .image_processing_beit import BeitImageProcessorKwargs
+
+
+# Adapted from transformers.models.beit.image_processing_beit.BeitImageProcessorKwargs
+class BeitImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    do_reduce_labels (`bool`, *optional*, defaults to `self.do_reduce_labels`):
+        Whether or not to reduce all label values of segmentation maps by 1. Usually used for datasets where 0
+        is used for background, and background itself is not included in all classes of a dataset (e.g.
+        ADE20k). The background label will be replaced by 255.
+    """
+
+    do_reduce_labels: bool
 
 
 @auto_docstring
-@requires(backends=("vision", "torch", "torchvision"))
 class BeitImageProcessorPil(PilBackend):
     """PIL backend for BEiT with reduce_label support."""
 
@@ -124,7 +131,7 @@ def _preprocess(
         images: list[np.ndarray],
         do_resize: bool,
         size: SizeDict,
-        resample: "PILImageResampling | tvF.InterpolationMode | int | None",
+        resample: PILImageResampling | None,
         do_center_crop: bool,
         crop_size: SizeDict,
         do_rescale: bool,
@@ -152,6 +159,7 @@ def _preprocess(
 
         return processed_images
 
+    @requires(backends=("torch",))
     def post_process_semantic_segmentation(self, outputs, target_sizes: list[tuple] | None = None):
         """
         Converts the output of [`BeitForSemanticSegmentation`] into semantic segmentation maps.
@@ -168,8 +176,8 @@ def post_process_semantic_segmentation(self, outputs, target_sizes: list[tuple]
             segmentation map of shape (height, width) corresponding to the target_sizes entry (if `target_sizes` is
             specified). Each entry of each `torch.Tensor` correspond to a semantic class id.
         """
-        if not is_torch_available():
-            raise ImportError("PyTorch is required for post_process_semantic_segmentation")
+        import torch
+        import torch.nn.functional as F
 
         logits = outputs.logits
 

src/transformers/models/bridgetower/image_processing_bridgetower.py

@@ -16,6 +16,8 @@
 from typing import Union
 
 import numpy as np
+import torch
+from torchvision.transforms.v2 import functional as tvF
 
 from ...image_processing_backends import TorchvisionBackend
 from ...image_processing_utils import BatchFeature
@@ -29,14 +31,7 @@
     get_image_size,
 )
 from ...processing_utils import ImagesKwargs, Unpack
-from ...utils import TensorType, auto_docstring, is_torch_available, is_torchvision_available
-
-
-if is_torch_available():
-    import torch
-
-if is_torchvision_available():
-    from torchvision.transforms.v2 import functional as tvF
+from ...utils import TensorType, auto_docstring, is_torch_available
 
 
 def get_resize_output_image_size(

src/transformers/models/bridgetower/image_processing_pil_bridgetower.py

@@ -20,16 +20,57 @@
 from ...image_utils import (
     OPENAI_CLIP_MEAN,
     OPENAI_CLIP_STD,
+    ChannelDimension,
     PILImageResampling,
     SizeDict,
+    get_image_size,
 )
-from ...processing_utils import Unpack
+from ...processing_utils import ImagesKwargs, Unpack
 from ...utils import TensorType, auto_docstring
-from ...utils.import_utils import requires
-from .image_processing_bridgetower import BridgeTowerImageProcessorKwargs, get_resize_output_image_size
 
 
-@requires(backends=("vision", "torch", "torchvision"))
+# Adapted from transformers.models.bridgetower.image_processing_bridgetower.BridgeTowerImageProcessorKwargs
+class BridgeTowerImageProcessorKwargs(ImagesKwargs, total=False):
+    r"""
+    size_divisor (`int`, *optional*, defaults to `self.size_divisor`):
+        The size by which to make sure both the height and width can be divided.
+    """
+
+    size_divisor: int
+
+
+# adapted from transformers.models.bridgetower.image_processing_bridgetower.get_resize_output_image_size
+def get_resize_output_image_size(
+    input_image: np.ndarray,
+    shorter: int = 800,
+    longer: int = 1333,
+    size_divisor: int = 32,
+) -> tuple[int, int]:
+    """Get output image size after resizing with size_divisor."""
+    input_height, input_width = get_image_size(input_image, channel_dim=ChannelDimension.FIRST)
+
+    min_size, max_size = shorter, longer
+    scale = min_size / min(input_height, input_width)
+
+    if input_height < input_width:
+        new_height = min_size
+        new_width = scale * input_width
+    else:
+        new_height = scale * input_height
+        new_width = min_size
+
+    if max(new_height, new_width) > max_size:
+        scale = max_size / max(new_height, new_width)
+        new_height = scale * new_height
+        new_width = scale * new_width
+
+    new_height, new_width = int(new_height + 0.5), int(new_width + 0.5)
+    new_height = new_height // size_divisor * size_divisor
+    new_width = new_width // size_divisor * size_divisor
+
+    return new_height, new_width
+
+
 @auto_docstring
 class BridgeTowerImageProcessorPil(PilBackend):
     """PIL backend for BridgeTower with custom resize and center_crop."""
@@ -57,7 +98,7 @@ def resize(
         self,
         image: np.ndarray,
         size: SizeDict,
-        resample: "PILImageResampling | int | None",
+        resample: "PILImageResampling | None",
         size_divisor: int = 32,
         **kwargs,
     ) -> np.ndarray:
@@ -82,7 +123,7 @@ def _preprocess(
         images: list[np.ndarray],
         do_resize: bool,
         size: SizeDict,
-        resample: "PILImageResampling | int | None",
+        resample: "PILImageResampling | None",
         do_center_crop: bool,
         crop_size: SizeDict,
         do_rescale: bool,

src/transformers/models/chameleon/image_processing_chameleon.py

@@ -15,6 +15,8 @@
 
 import numpy as np
 import PIL.Image
+import torch
+from torchvision.transforms.v2 import functional as tvF
 
 from ...image_processing_backends import TorchvisionBackend
 from ...image_utils import (
@@ -23,15 +25,9 @@
     SizeDict,
 )
 from ...processing_utils import ImagesKwargs, Unpack
-from ...utils import auto_docstring, is_torch_available, is_torchvision_available, logging
+from ...utils import auto_docstring, logging
 
 
-if is_torch_available():
-    import torch
-
-if is_torchvision_available():
-    from torchvision.transforms.v2 import functional as tvF
-
 logger = logging.get_logger(__name__)
 
 

src/transformers/models/chmv2/image_processing_chmv2.py

@@ -23,22 +23,16 @@
 from typing import Union
 
 import torch
+import torch.nn.functional as F
+from torchvision.transforms.v2 import functional as tvF
 
 from ...image_processing_backends import TorchvisionBackend
 from ...image_processing_base import BatchFeature
 from ...image_transforms import group_images_by_shape, reorder_images
 from ...image_utils import ChannelDimension, ImageInput, PILImageResampling, SizeDict
 from ...modeling_outputs import DepthEstimatorOutput
 from ...processing_utils import ImagesKwargs, Unpack
-from ...utils import TensorType, auto_docstring, is_torch_available, is_torchvision_available, requires_backends
-
-
-if is_torchvision_available():
-    import torchvision.transforms.v2.functional as tvF
-
-
-if is_torch_available():
-    import torch.nn.functional as F
+from ...utils import TensorType, auto_docstring, is_torch_available, requires_backends
 
 
 class CHMv2ImageProcessorKwargs(ImagesKwargs, total=False):