Fix hidden torchvision>=0.15 dependency issue (#39928)

* use pil_torch_interpolation_mapping for NEAREST/NEAREST_EXACT

* fix min torchvision version

* use InterpolationMode directly

* remove unused is_torchvision_greater_or_equal,

* nit

src/transformers/image_utils.py

@@ -31,6 +31,7 @@ from .utils import (
     is_torch_available,
     is_torch_tensor,
     is_torchvision_available,
+    is_torchvision_v2_available,
     is_vision_available,
     logging,
     requires_backends,
@@ -59,7 +60,9 @@ if is_vision_available():
         from torchvision.transforms import InterpolationMode
 
         pil_torch_interpolation_mapping = {
-            PILImageResampling.NEAREST: InterpolationMode.NEAREST_EXACT,
+            PILImageResampling.NEAREST: InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else InterpolationMode.NEAREST,
             PILImageResampling.BOX: InterpolationMode.BOX,
             PILImageResampling.BILINEAR: InterpolationMode.BILINEAR,
             PILImageResampling.HAMMING: InterpolationMode.HAMMING,

src/transformers/models/conditional_detr/image_processing_conditional_detr_fast.py

@@ -57,6 +57,7 @@ if is_torch_available():
 if is_torchvision_v2_available():
     from torchvision.io import read_image
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.io import read_image
     from torchvision.transforms import functional as F
@@ -454,10 +455,16 @@ class ConditionalDetrImageProcessorFast(BaseImageProcessorFast):
                 The target size of the image, as returned by the preprocessing `resize` step.
             threshold (`float`, *optional*, defaults to 0.5):
                 The threshold used to binarize the segmentation masks.
-            resample (`InterpolationMode`, defaults to `InterpolationMode.NEAREST_EXACT`):
+            resample (`InterpolationMode`, defaults to `F.InterpolationMode.NEAREST_EXACT`):
                 The resampling filter to use when resizing the masks.
         """
-        interpolation = interpolation if interpolation is not None else F.InterpolationMode.NEAREST_EXACT
+        interpolation = (
+            interpolation
+            if interpolation is not None
+            else F.InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else F.InterpolationMode.NEAREST
+        )
         ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
 
         new_annotation = {}

src/transformers/models/deformable_detr/image_processing_deformable_detr_fast.py

@@ -48,6 +48,7 @@ if is_torch_available():
 if is_torchvision_v2_available():
     from torchvision.io import read_image
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.io import read_image
     from torchvision.transforms import functional as F
@@ -445,10 +446,16 @@ class DeformableDetrImageProcessorFast(BaseImageProcessorFast):
                 The target size of the image, as returned by the preprocessing `resize` step.
             threshold (`float`, *optional*, defaults to 0.5):
                 The threshold used to binarize the segmentation masks.
-            resample (`InterpolationMode`, defaults to `InterpolationMode.NEAREST_EXACT`):
+            resample (`InterpolationMode`, defaults to `F.InterpolationMode.NEAREST_EXACT`):
                 The resampling filter to use when resizing the masks.
         """
-        interpolation = interpolation if interpolation is not None else F.InterpolationMode.NEAREST_EXACT
+        interpolation = (
+            interpolation
+            if interpolation is not None
+            else F.InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else F.InterpolationMode.NEAREST
+        )
         ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
 
         new_annotation = {}

src/transformers/models/detr/image_processing_detr_fast.py

@@ -70,6 +70,7 @@ if is_vision_available():
 if is_torchvision_v2_available():
     from torchvision.io import read_image
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.io import read_image
     from torchvision.transforms import functional as F
@@ -466,10 +467,16 @@ class DetrImageProcessorFast(BaseImageProcessorFast):
                 The target size of the image, as returned by the preprocessing `resize` step.
             threshold (`float`, *optional*, defaults to 0.5):
                 The threshold used to binarize the segmentation masks.
-            resample (`InterpolationMode`, defaults to `InterpolationMode.NEAREST_EXACT`):
+            resample (`InterpolationMode`, defaults to `F.InterpolationMode.NEAREST_EXACT`):
                 The resampling filter to use when resizing the masks.
         """
-        interpolation = interpolation if interpolation is not None else F.InterpolationMode.NEAREST_EXACT
+        interpolation = (
+            interpolation
+            if interpolation is not None
+            else F.InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else F.InterpolationMode.NEAREST
+        )
         ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
 
         new_annotation = {}

src/transformers/models/eomt/image_processing_eomt_fast.py

@@ -33,7 +33,6 @@ from ...image_utils import (
     ImageInput,
     PILImageResampling,
     SizeDict,
-    pil_torch_interpolation_mapping,
 )
 from ...processing_utils import Unpack
 from ...utils import (
@@ -210,7 +209,9 @@ class EomtImageProcessorFast(BaseImageProcessorFast):
                     "do_normalize": False,
                     "do_rescale": False,
                     # Nearest interpolation is used for segmentation maps instead of BILINEAR.
-                    "interpolation": pil_torch_interpolation_mapping[PILImageResampling.NEAREST],
+                    "interpolation": F.InterpolationMode.NEAREST_EXACT
+                    if is_torchvision_v2_available()
+                    else F.InterpolationMode.NEAREST,
                 }
             )
 

src/transformers/models/grounding_dino/image_processing_grounding_dino_fast.py

@@ -51,6 +51,7 @@ if is_torch_available():
 if is_torchvision_v2_available():
     from torchvision.io import read_image
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.io import read_image
     from torchvision.transforms import functional as F
@@ -476,10 +477,16 @@ class GroundingDinoImageProcessorFast(BaseImageProcessorFast):
                 The target size of the image, as returned by the preprocessing `resize` step.
             threshold (`float`, *optional*, defaults to 0.5):
                 The threshold used to binarize the segmentation masks.
-            resample (`InterpolationMode`, defaults to `InterpolationMode.NEAREST_EXACT`):
+            resample (`InterpolationMode`, defaults to `F.InterpolationMode.NEAREST_EXACT`):
                 The resampling filter to use when resizing the masks.
         """
-        interpolation = interpolation if interpolation is not None else F.InterpolationMode.NEAREST_EXACT
+        interpolation = (
+            interpolation
+            if interpolation is not None
+            else F.InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else F.InterpolationMode.NEAREST
+        )
         ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
 
         new_annotation = {}

src/transformers/models/mask2former/image_processing_mask2former_fast.py

@@ -62,6 +62,7 @@ if is_torch_available():
 
 if is_torchvision_v2_available():
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.transforms import functional as F
 
@@ -369,7 +370,9 @@ class Mask2FormerImageProcessorFast(BaseImageProcessorFast):
                         image=grouped_segmentation_maps[shape],
                         size=size,
                         size_divisor=size_divisor,
-                        interpolation=F.InterpolationMode.NEAREST_EXACT,
+                        interpolation=F.InterpolationMode.NEAREST_EXACT
+                        if is_torchvision_v2_available()
+                        else F.InterpolationMode.NEAREST,
                     )
             resized_images_grouped[shape] = stacked_images
             if segmentation_maps is not None:

src/transformers/models/maskformer/image_processing_maskformer_fast.py

@@ -66,6 +66,7 @@ if is_torch_available():
 
 if is_torchvision_v2_available():
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.transforms import functional as F
 
@@ -370,7 +371,9 @@ class MaskFormerImageProcessorFast(BaseImageProcessorFast):
                         image=grouped_segmentation_maps[shape],
                         size=size,
                         size_divisor=size_divisor,
-                        interpolation=F.InterpolationMode.NEAREST_EXACT,
+                        interpolation=F.InterpolationMode.NEAREST_EXACT
+                        if is_torchvision_v2_available()
+                        else F.InterpolationMode.NEAREST,
                     )
             resized_images_grouped[shape] = stacked_images
             if segmentation_maps is not None:

src/transformers/models/mobilenet_v2/image_processing_mobilenet_v2_fast.py

@@ -31,7 +31,6 @@ from ...image_utils import (
     PILImageResampling,
     SizeDict,
     is_torch_tensor,
-    pil_torch_interpolation_mapping,
 )
 from ...processing_utils import Unpack
 from ...utils import (
@@ -139,7 +138,9 @@ class MobileNetV2ImageProcessorFast(BaseImageProcessorFast):
                     "do_normalize": False,
                     "do_rescale": False,
                     # Nearest interpolation is used for segmentation maps instead of BILINEAR.
-                    "interpolation": pil_torch_interpolation_mapping[PILImageResampling.NEAREST],
+                    "interpolation": F.InterpolationMode.NEAREST_EXACT
+                    if is_torchvision_v2_available()
+                    else F.InterpolationMode.NEAREST,
                 }
             )
 

src/transformers/models/mobilevit/image_processing_mobilevit_fast.py

@@ -29,7 +29,6 @@ from ...image_utils import (
     PILImageResampling,
     SizeDict,
     is_torch_tensor,
-    pil_torch_interpolation_mapping,
 )
 from ...processing_utils import Unpack
 from ...utils import (
@@ -140,7 +139,9 @@ class MobileViTImageProcessorFast(BaseImageProcessorFast):
                     "do_rescale": False,
                     "do_flip_channel_order": False,
                     # Nearest interpolation is used for segmentation maps instead of BILINEAR.
-                    "interpolation": pil_torch_interpolation_mapping[PILImageResampling.NEAREST],
+                    "interpolation": F.InterpolationMode.NEAREST_EXACT
+                    if is_torchvision_v2_available()
+                    else F.InterpolationMode.NEAREST,
                 }
             )
 

src/transformers/models/oneformer/image_processing_oneformer_fast.py

@@ -457,7 +457,11 @@ class OneFormerImageProcessorFast(BaseImageProcessorFast):
             for shape, stacked_segmentation_maps in grouped_segmentation_maps.items():
                 if do_resize:
                     stacked_segmentation_maps = self.resize(
-                        stacked_segmentation_maps, size=size, interpolation=F.InterpolationMode.NEAREST_EXACT
+                        stacked_segmentation_maps,
+                        size=size,
+                        interpolation=F.InterpolationMode.NEAREST_EXACT
+                        if is_torchvision_v2_available()
+                        else F.InterpolationMode.NEAREST,
                     )
                 processed_segmentation_maps_grouped[shape] = stacked_segmentation_maps
             processed_segmentation_maps = reorder_images(

src/transformers/models/rt_detr/image_processing_rt_detr_fast.py

@@ -264,10 +264,16 @@ class RTDetrImageProcessorFast(BaseImageProcessorFast):
                 The target size of the image, as returned by the preprocessing `resize` step.
             threshold (`float`, *optional*, defaults to 0.5):
                 The threshold used to binarize the segmentation masks.
-            resample (`InterpolationMode`, defaults to `InterpolationMode.NEAREST_EXACT`):
+            resample (`InterpolationMode`, defaults to `F.InterpolationMode.NEAREST_EXACT`):
                 The resampling filter to use when resizing the masks.
         """
-        interpolation = interpolation if interpolation is not None else F.InterpolationMode.NEAREST_EXACT
+        interpolation = (
+            interpolation
+            if interpolation is not None
+            else F.InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else F.InterpolationMode.NEAREST
+        )
         ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
 
         new_annotation = {}

src/transformers/models/sam/image_processing_sam_fast.py

@@ -36,7 +36,6 @@ from ...image_utils import (
     ImageInput,
     PILImageResampling,
     SizeDict,
-    pil_torch_interpolation_mapping,
 )
 from ...processing_utils import Unpack
 from ...utils import (
@@ -244,7 +243,9 @@ class SamImageProcessorFast(BaseImageProcessorFast):
                 {
                     "do_normalize": False,
                     "do_rescale": False,
-                    "interpolation": pil_torch_interpolation_mapping[PILImageResampling.NEAREST],
+                    "interpolation": F.InterpolationMode.NEAREST_EXACT
+                    if is_torchvision_v2_available()
+                    else F.InterpolationMode.NEAREST,
                     "size": segmentation_maps_kwargs.pop("mask_size"),
                     "pad_size": segmentation_maps_kwargs.pop("mask_pad_size"),
                 }

src/transformers/models/segformer/image_processing_segformer_fast.py

@@ -36,7 +36,6 @@ from ...image_utils import (
     PILImageResampling,
     SizeDict,
     is_torch_tensor,
-    pil_torch_interpolation_mapping,
 )
 from ...processing_utils import Unpack
 from ...utils import (
@@ -143,7 +142,9 @@ class SegformerImageProcessorFast(BaseImageProcessorFast):
                     "do_normalize": False,
                     "do_rescale": False,
                     # Nearest interpolation is used for segmentation maps instead of BILINEAR.
-                    "interpolation": pil_torch_interpolation_mapping[PILImageResampling.NEAREST],
+                    "interpolation": F.InterpolationMode.NEAREST_EXACT
+                    if is_torchvision_v2_available()
+                    else F.InterpolationMode.NEAREST,
                 }
             )
             processed_segmentation_maps = self._preprocess(

src/transformers/models/segformer/modular_segformer.py

@@ -30,7 +30,6 @@ from ...image_utils import (
     ImageInput,
     PILImageResampling,
     SizeDict,
-    pil_torch_interpolation_mapping,
 )
 from ...processing_utils import Unpack
 from ...utils import (
@@ -100,7 +99,9 @@ class SegformerImageProcessorFast(BeitImageProcessorFast):
                     "do_normalize": False,
                     "do_rescale": False,
                     # Nearest interpolation is used for segmentation maps instead of BILINEAR.
-                    "interpolation": pil_torch_interpolation_mapping[PILImageResampling.NEAREST],
+                    "interpolation": F.InterpolationMode.NEAREST_EXACT
+                    if is_torchvision_v2_available()
+                    else F.InterpolationMode.NEAREST,
                 }
             )
             processed_segmentation_maps = self._preprocess(

src/transformers/models/yolos/image_processing_yolos_fast.py

@@ -47,6 +47,7 @@ if is_torch_available():
 if is_torchvision_v2_available():
     from torchvision.io import read_image
     from torchvision.transforms.v2 import functional as F
+
 elif is_torchvision_available():
     from torchvision.io import read_image
     from torchvision.transforms import functional as F
@@ -493,10 +494,16 @@ class YolosImageProcessorFast(BaseImageProcessorFast):
                 The target size of the image, as returned by the preprocessing `resize` step.
             threshold (`float`, *optional*, defaults to 0.5):
                 The threshold used to binarize the segmentation masks.
-            resample (`InterpolationMode`, defaults to `InterpolationMode.NEAREST_EXACT`):
+            resample (`InterpolationMode`, defaults to `F.InterpolationMode.NEAREST_EXACT`):
                 The resampling filter to use when resizing the masks.
         """
-        interpolation = interpolation if interpolation is not None else F.InterpolationMode.NEAREST_EXACT
+        interpolation = (
+            interpolation
+            if interpolation is not None
+            else F.InterpolationMode.NEAREST_EXACT
+            if is_torchvision_v2_available()
+            else F.InterpolationMode.NEAREST
+        )
         ratio_height, ratio_width = [target / orig for target, orig in zip(target_size, orig_size)]
 
         new_annotation = {}