Add owlv2 fast processor (#39041)

* add owlv2 fast image processor

* add Owlv2ImageProcessorFast to Owlv2Processor image_processor_class

* add Owlv2ImageProcessorFast to Owlv2Processor image_processor_class

* change references to owlVit to owlv2 in docstrings for post process methods

* change type hints from List, Dict, Tuple to list, dict, tuple

* remove unused typing imports

* add disable grouping argument to group images by shape

* run make quality and repo-consistency

* use modular

* fix auto_docstring

---------

Co-authored-by: Lewis Marshall <lewism@elderda.co.uk>
Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co>

docs/source/en/model_doc/owlv2.md

@@ -106,6 +106,13 @@ Usage of OWLv2 is identical to [OWL-ViT](owlvit) with a new, updated image proce
     - post_process_object_detection
     - post_process_image_guided_detection
 
+## Owlv2ImageProcessorFast
+
+[[autodoc]] Owlv2ImageProcessorFast
+    - preprocess
+    - post_process_object_detection
+    - post_process_image_guided_detection
+
 ## Owlv2Processor
 
 [[autodoc]] Owlv2Processor

src/transformers/models/auto/image_processing_auto.py

@@ -131,7 +131,7 @@ else:
             ("nat", ("ViTImageProcessor", "ViTImageProcessorFast")),
             ("nougat", ("NougatImageProcessor", "NougatImageProcessorFast")),
             ("oneformer", ("OneFormerImageProcessor", "OneFormerImageProcessorFast")),
-            ("owlv2", ("Owlv2ImageProcessor",)),
+            ("owlv2", ("Owlv2ImageProcessor", "Owlv2ImageProcessorFast")),
             ("owlvit", ("OwlViTImageProcessor", "OwlViTImageProcessorFast")),
             ("paligemma", ("SiglipImageProcessor", "SiglipImageProcessorFast")),
             ("perceiver", ("PerceiverImageProcessor", "PerceiverImageProcessorFast")),

src/transformers/models/owlv2/__init__.py

@@ -20,6 +20,7 @@ from ...utils.import_utils import define_import_structure
 if TYPE_CHECKING:
     from .configuration_owlv2 import *
     from .image_processing_owlv2 import *
+    from .image_processing_owlv2_fast import *
     from .modeling_owlv2 import *
     from .processing_owlv2 import *
 else:

src/transformers/models/owlv2/image_processing_owlv2_fast.py

@@ -0,0 +1,427 @@
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+#           This file was automatically generated from src/transformers/models/owlv2/modular_owlv2.py.
+#               Do NOT edit this file manually as any edits will be overwritten by the generation of
+#             the file from the modular. If any change should be done, please apply the change to the
+#                          modular_owlv2.py file directly. One of our CI enforces this.
+#                🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨🚨
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team. 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 warnings
+from typing import TYPE_CHECKING, Optional, Union
+
+from ...image_processing_utils_fast import BaseImageProcessorFast, BatchFeature, DefaultFastImageProcessorKwargs
+from ...image_transforms import center_to_corners_format, group_images_by_shape, reorder_images
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+)
+from ...processing_utils import Unpack
+from ...utils import (
+    TensorType,
+    auto_docstring,
+    is_torch_available,
+    is_torchvision_available,
+    is_torchvision_v2_available,
+)
+
+
+if is_torch_available():
+    import torch
+
+
+if is_torchvision_v2_available():
+    from torchvision.transforms.v2 import functional as F
+elif is_torchvision_available():
+    from torchvision.transforms import functional as F
+
+
+if TYPE_CHECKING:
+    from .modeling_owlv2 import Owlv2ObjectDetectionOutput
+
+
+if is_torch_available():
+    from .image_processing_owlv2 import _scale_boxes, box_iou
+
+
+class Owlv2FastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+    r"""
+    do_pad (`bool`, *optional*, defaults to `True`):
+        Controls whether to pad the image. Can be overridden by the `do_pad` parameter in the `preprocess`
+        method. If `True`, padding will be applied to the bottom and right of the image with grey pixels.
+    """
+
+    do_pad: Optional[bool]
+
+
+@auto_docstring
+class Owlv2ImageProcessorFast(BaseImageProcessorFast):
+    resample = PILImageResampling.BILINEAR
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 960, "width": 960}
+    default_to_square = True
+    crop_size = None
+    do_resize = True
+    do_center_crop = None
+    do_rescale = True
+    do_normalize = True
+    do_convert_rgb = None
+    model_input_names = ["pixel_values"]
+    rescale_factor = 1 / 255
+    do_pad = True
+    valid_kwargs = Owlv2FastImageProcessorKwargs
+
+    def post_process(self, outputs, target_sizes):
+        """
+        Converts the raw output of [`Owlv2ForObjectDetection`] into final bounding boxes in (top_left_x, top_left_y,
+        bottom_right_x, bottom_right_y) format.
+
+        Args:
+            outputs ([`Owlv2ObjectDetectionOutput`]):
+                Raw outputs of the model.
+            target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
+                Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
+                image size (before any data augmentation). For visualization, this should be the image size after data
+                augment, but before padding.
+        Returns:
+            `list[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model.
+        """
+        # TODO: (amy) add support for other frameworks
+        warnings.warn(
+            "`post_process` is deprecated and will be removed in v5 of Transformers, please use"
+            " `post_process_object_detection` instead, with `threshold=0.` for equivalent results.",
+            FutureWarning,
+        )
+
+        logits, boxes = outputs.logits, outputs.pred_boxes
+
+        if len(logits) != len(target_sizes):
+            raise ValueError("Make sure that you pass in as many target sizes as the batch dimension of the logits")
+        if target_sizes.shape[1] != 2:
+            raise ValueError("Each element of target_sizes must contain the size (h, w) of each image of the batch")
+
+        probs = torch.max(logits, dim=-1)
+        scores = torch.sigmoid(probs.values)
+        labels = probs.indices
+
+        # Convert to [x0, y0, x1, y1] format
+        boxes = center_to_corners_format(boxes)
+
+        # Convert from relative [0, 1] to absolute [0, height] coordinates
+        img_h, img_w = target_sizes.unbind(1)
+        scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1).to(boxes.device)
+        boxes = boxes * scale_fct[:, None, :]
+
+        results = [{"scores": s, "labels": l, "boxes": b} for s, l, b in zip(scores, labels, boxes)]
+
+        return results
+
+    def post_process_object_detection(
+        self,
+        outputs: "Owlv2ObjectDetectionOutput",
+        threshold: float = 0.1,
+        target_sizes: Optional[Union[TensorType, list[tuple]]] = None,
+    ):
+        """
+        Converts the raw output of [`Owlv2ForObjectDetection`] into final bounding boxes in (top_left_x, top_left_y,
+        bottom_right_x, bottom_right_y) format.
+
+        Args:
+            outputs ([`Owlv2ObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*, defaults to 0.1):
+                Score threshold to keep object detection predictions.
+            target_sizes (`torch.Tensor` or `list[tuple[int, int]]`, *optional*):
+                Tensor of shape `(batch_size, 2)` or list of tuples (`tuple[int, int]`) containing the target size
+                `(height, width)` of each image in the batch. If unset, predictions will not be resized.
+
+        Returns:
+            `list[Dict]`: A list of dictionaries, each dictionary containing the following keys:
+            - "scores": The confidence scores for each predicted box on the image.
+            - "labels": Indexes of the classes predicted by the model on the image.
+            - "boxes": Image bounding boxes in (top_left_x, top_left_y, bottom_right_x, bottom_right_y) format.
+        """
+        batch_logits, batch_boxes = outputs.logits, outputs.pred_boxes
+        batch_size = len(batch_logits)
+
+        if target_sizes is not None and len(target_sizes) != batch_size:
+            raise ValueError("Make sure that you pass in as many target sizes as images")
+
+        # batch_logits of shape (batch_size, num_queries, num_classes)
+        batch_class_logits = torch.max(batch_logits, dim=-1)
+        batch_scores = torch.sigmoid(batch_class_logits.values)
+        batch_labels = batch_class_logits.indices
+
+        # Convert to [x0, y0, x1, y1] format
+        batch_boxes = center_to_corners_format(batch_boxes)
+
+        # Convert from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            batch_boxes = _scale_boxes(batch_boxes, target_sizes)
+
+        results = []
+        for scores, labels, boxes in zip(batch_scores, batch_labels, batch_boxes):
+            keep = scores > threshold
+            scores = scores[keep]
+            labels = labels[keep]
+            boxes = boxes[keep]
+            results.append({"scores": scores, "labels": labels, "boxes": boxes})
+
+        return results
+
+    def post_process_image_guided_detection(self, outputs, threshold=0.0, nms_threshold=0.3, target_sizes=None):
+        """
+        Converts the output of [`Owlv2ForObjectDetection.image_guided_detection`] into the format expected by the COCO
+        api.
+
+        Args:
+            outputs ([`Owlv2ImageGuidedObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*, defaults to 0.0):
+                Minimum confidence threshold to use to filter out predicted boxes.
+            nms_threshold (`float`, *optional*, defaults to 0.3):
+                IoU threshold for non-maximum suppression of overlapping boxes.
+            target_sizes (`torch.Tensor`, *optional*):
+                Tensor of shape (batch_size, 2) where each entry is the (height, width) of the corresponding image in
+                the batch. If set, predicted normalized bounding boxes are rescaled to the target sizes. If left to
+                None, predictions will not be unnormalized.
+
+        Returns:
+            `list[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
+            in the batch as predicted by the model. All labels are set to None as
+            `Owlv2ForObjectDetection.image_guided_detection` perform one-shot object detection.
+        """
+        logits, target_boxes = outputs.logits, outputs.target_pred_boxes
+
+        if target_sizes is not None and len(logits) != len(target_sizes):
+            raise ValueError("Make sure that you pass in as many target sizes as the batch dimension of the logits")
+        if target_sizes is not None and target_sizes.shape[1] != 2:
+            raise ValueError("Each element of target_sizes must contain the size (h, w) of each image of the batch")
+
+        probs = torch.max(logits, dim=-1)
+        scores = torch.sigmoid(probs.values)
+
+        # Convert to [x0, y0, x1, y1] format
+        target_boxes = center_to_corners_format(target_boxes)
+
+        # Apply non-maximum suppression (NMS)
+        if nms_threshold < 1.0:
+            for idx in range(target_boxes.shape[0]):
+                for i in torch.argsort(-scores[idx]):
+                    if not scores[idx][i]:
+                        continue
+
+                    ious = box_iou(target_boxes[idx][i, :].unsqueeze(0), target_boxes[idx])[0][0]
+                    ious[i] = -1.0  # Mask self-IoU.
+                    scores[idx][ious > nms_threshold] = 0.0
+
+        # Convert from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            target_boxes = _scale_boxes(target_boxes, target_sizes)
+
+        # Compute box display alphas based on prediction scores
+        results = []
+        alphas = torch.zeros_like(scores)
+
+        for idx in range(target_boxes.shape[0]):
+            # Select scores for boxes matching the current query:
+            query_scores = scores[idx]
+            if not query_scores.nonzero().numel():
+                continue
+
+            # Apply threshold on scores before scaling
+            query_scores[query_scores < threshold] = 0.0
+
+            # Scale box alpha such that the best box for each query has alpha 1.0 and the worst box has alpha 0.1.
+            # All other boxes will either belong to a different query, or will not be shown.
+            max_score = torch.max(query_scores) + 1e-6
+            query_alphas = (query_scores - (max_score * 0.1)) / (max_score * 0.9)
+            query_alphas = torch.clip(query_alphas, 0.0, 1.0)
+            alphas[idx] = query_alphas
+
+            mask = alphas[idx] > 0
+            box_scores = alphas[idx][mask]
+            boxes = target_boxes[idx][mask]
+            results.append({"scores": box_scores, "labels": None, "boxes": boxes})
+
+        return results
+
+    def __init__(self, **kwargs: Unpack[Owlv2FastImageProcessorKwargs]):
+        super().__init__(**kwargs)
+
+    @auto_docstring
+    def preprocess(self, images: ImageInput, **kwargs: Unpack[Owlv2FastImageProcessorKwargs]):
+        return super().preprocess(images, **kwargs)
+
+    def _pad_images(self, images: "torch.Tensor", constant_value: float = 0.5) -> "torch.Tensor":
+        """
+        Pad an image with zeros to the given size.
+        """
+        height, width = images.shape[-2:]
+        size = max(height, width)
+        pad_bottom = size - height
+        pad_right = size - width
+
+        padding = (0, 0, pad_right, pad_bottom)
+        padded_image = F.pad(images, padding, fill=constant_value)
+        return padded_image
+
+    def pad(
+        self,
+        images: list["torch.Tensor"],
+        disable_grouping: Optional[bool],
+        constant_value: float = 0.5,
+    ) -> list["torch.Tensor"]:
+        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            stacked_images = self._pad_images(
+                stacked_images,
+                constant_value=constant_value,
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+
+        return processed_images
+
+    def resize(
+        self,
+        image: "torch.Tensor",
+        size: SizeDict,
+        anti_aliasing: bool = True,
+        anti_aliasing_sigma=None,
+        **kwargs,
+    ) -> "torch.Tensor":
+        """
+        Resize an image as per the original implementation.
+
+        Args:
+            image (`Tensor`):
+                Image to resize.
+            size (`dict[str, int]`):
+                Dictionary containing the height and width to resize the image to.
+            anti_aliasing (`bool`, *optional*, defaults to `True`):
+                Whether to apply anti-aliasing when downsampling the image.
+            anti_aliasing_sigma (`float`, *optional*, defaults to `None`):
+                Standard deviation for Gaussian kernel when downsampling the image. If `None`, it will be calculated
+                automatically.
+        """
+        output_shape = (size.height, size.width)
+
+        input_shape = image.shape
+
+        # select height and width from input tensor
+        factors = torch.tensor(input_shape[2:]).to(image.device) / torch.tensor(output_shape).to(image.device)
+
+        if anti_aliasing:
+            if anti_aliasing_sigma is None:
+                anti_aliasing_sigma = ((factors - 1) / 2).clamp(min=0)
+            else:
+                anti_aliasing_sigma = torch.atleast_1d(anti_aliasing_sigma) * torch.ones_like(factors)
+                if torch.any(anti_aliasing_sigma < 0):
+                    raise ValueError("Anti-aliasing standard deviation must be greater than or equal to zero")
+                elif torch.any((anti_aliasing_sigma > 0) & (factors <= 1)):
+                    warnings.warn(
+                        "Anti-aliasing standard deviation greater than zero but not down-sampling along all axes"
+                    )
+            if torch.any(anti_aliasing_sigma == 0):
+                filtered = image
+            else:
+                kernel_sizes = 2 * torch.ceil(3 * anti_aliasing_sigma).int() + 1
+
+                filtered = F.gaussian_blur(
+                    image, (kernel_sizes[0], kernel_sizes[1]), sigma=anti_aliasing_sigma.tolist()
+                )
+
+        else:
+            filtered = image
+
+        out = F.resize(filtered, size=(size.height, size.width), antialias=False)
+
+        return out
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        size: SizeDict,
+        interpolation: Optional["F.InterpolationMode"],
+        do_pad: bool,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: Optional[Union[float, list[float]]],
+        image_std: Optional[Union[float, list[float]]],
+        disable_grouping: Optional[bool],
+        return_tensors: Optional[Union[str, TensorType]],
+        **kwargs,
+    ) -> BatchFeature:
+        # Group images by size for batched resizing
+        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+
+        for shape, stacked_images in grouped_images.items():
+            # Rescale images before other operations as done in original implementation
+            stacked_images = self.rescale_and_normalize(
+                stacked_images, do_rescale, rescale_factor, False, image_mean, image_std
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+
+        if do_pad:
+            processed_images = self.pad(processed_images, disable_grouping=disable_grouping)
+
+        grouped_images, grouped_images_index = group_images_by_shape(
+            processed_images, disable_grouping=disable_grouping
+        )
+        resized_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            if do_resize:
+                resized_stack = self.resize(
+                    image=stacked_images,
+                    size=size,
+                    interpolation=interpolation,
+                    input_data_format=ChannelDimension.FIRST,
+                )
+                resized_images_grouped[shape] = resized_stack
+        resized_images = reorder_images(resized_images_grouped, grouped_images_index)
+
+        # Group images by size for further processing
+        # Needed in case do_resize is False, or resize returns images with different sizes
+        grouped_images, grouped_images_index = group_images_by_shape(resized_images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            # Fused rescale and normalize
+            stacked_images = self.rescale_and_normalize(
+                stacked_images, False, rescale_factor, do_normalize, image_mean, image_std
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+
+        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
+
+        return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
+
+
+__all__ = ["Owlv2ImageProcessorFast"]

src/transformers/models/owlv2/modular_owlv2.py

@@ -0,0 +1,240 @@
+# coding=utf-8
+# Copyright 2025 The HuggingFace Inc. team. 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.
+"""Fast Image processor class for OWLv2."""
+
+import warnings
+from typing import Optional, Union
+
+from transformers.models.owlvit.image_processing_owlvit_fast import OwlViTImageProcessorFast
+
+from ...image_processing_utils_fast import (
+    BatchFeature,
+    DefaultFastImageProcessorKwargs,
+)
+from ...image_transforms import group_images_by_shape, reorder_images
+from ...image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    SizeDict,
+)
+from ...processing_utils import Unpack
+from ...utils import (
+    TensorType,
+    auto_docstring,
+    is_torch_available,
+    is_torchvision_available,
+    is_torchvision_v2_available,
+)
+
+
+if is_torch_available():
+    import torch
+
+
+if is_torchvision_v2_available():
+    from torchvision.transforms.v2 import functional as F
+elif is_torchvision_available():
+    from torchvision.transforms import functional as F
+
+
+class Owlv2FastImageProcessorKwargs(DefaultFastImageProcessorKwargs):
+    r"""
+    do_pad (`bool`, *optional*, defaults to `True`):
+        Controls whether to pad the image. Can be overridden by the `do_pad` parameter in the `preprocess`
+        method. If `True`, padding will be applied to the bottom and right of the image with grey pixels.
+    """
+
+    do_pad: Optional[bool]
+
+
+@auto_docstring
+class Owlv2ImageProcessorFast(OwlViTImageProcessorFast):
+    resample = PILImageResampling.BILINEAR
+    image_mean = OPENAI_CLIP_MEAN
+    image_std = OPENAI_CLIP_STD
+    size = {"height": 960, "width": 960}
+    rescale_factor = 1 / 255
+    do_resize = True
+    do_rescale = True
+    do_normalize = True
+    do_pad = True
+    valid_kwargs = Owlv2FastImageProcessorKwargs
+    crop_size = None
+    do_center_crop = None
+
+    def __init__(self, **kwargs: Unpack[Owlv2FastImageProcessorKwargs]):
+        OwlViTImageProcessorFast().__init__(**kwargs)
+
+    @auto_docstring
+    def preprocess(self, images: ImageInput, **kwargs: Unpack[Owlv2FastImageProcessorKwargs]):
+        return OwlViTImageProcessorFast().preprocess(images, **kwargs)
+
+    def _pad_images(self, images: "torch.Tensor", constant_value: float = 0.5) -> "torch.Tensor":
+        """
+        Pad an image with zeros to the given size.
+        """
+        height, width = images.shape[-2:]
+        size = max(height, width)
+        pad_bottom = size - height
+        pad_right = size - width
+
+        padding = (0, 0, pad_right, pad_bottom)
+        padded_image = F.pad(images, padding, fill=constant_value)
+        return padded_image
+
+    def pad(
+        self,
+        images: list["torch.Tensor"],
+        disable_grouping: Optional[bool],
+        constant_value: float = 0.5,
+    ) -> list["torch.Tensor"]:
+        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            stacked_images = self._pad_images(
+                stacked_images,
+                constant_value=constant_value,
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+
+        return processed_images
+
+    def resize(
+        self,
+        image: "torch.Tensor",
+        size: SizeDict,
+        anti_aliasing: bool = True,
+        anti_aliasing_sigma=None,
+        **kwargs,
+    ) -> "torch.Tensor":
+        """
+        Resize an image as per the original implementation.
+
+        Args:
+            image (`Tensor`):
+                Image to resize.
+            size (`dict[str, int]`):
+                Dictionary containing the height and width to resize the image to.
+            anti_aliasing (`bool`, *optional*, defaults to `True`):
+                Whether to apply anti-aliasing when downsampling the image.
+            anti_aliasing_sigma (`float`, *optional*, defaults to `None`):
+                Standard deviation for Gaussian kernel when downsampling the image. If `None`, it will be calculated
+                automatically.
+        """
+        output_shape = (size.height, size.width)
+
+        input_shape = image.shape
+
+        # select height and width from input tensor
+        factors = torch.tensor(input_shape[2:]).to(image.device) / torch.tensor(output_shape).to(image.device)
+
+        if anti_aliasing:
+            if anti_aliasing_sigma is None:
+                anti_aliasing_sigma = ((factors - 1) / 2).clamp(min=0)
+            else:
+                anti_aliasing_sigma = torch.atleast_1d(anti_aliasing_sigma) * torch.ones_like(factors)
+                if torch.any(anti_aliasing_sigma < 0):
+                    raise ValueError("Anti-aliasing standard deviation must be greater than or equal to zero")
+                elif torch.any((anti_aliasing_sigma > 0) & (factors <= 1)):
+                    warnings.warn(
+                        "Anti-aliasing standard deviation greater than zero but not down-sampling along all axes"
+                    )
+            if torch.any(anti_aliasing_sigma == 0):
+                filtered = image
+            else:
+                kernel_sizes = 2 * torch.ceil(3 * anti_aliasing_sigma).int() + 1
+
+                filtered = F.gaussian_blur(
+                    image, (kernel_sizes[0], kernel_sizes[1]), sigma=anti_aliasing_sigma.tolist()
+                )
+
+        else:
+            filtered = image
+
+        out = F.resize(filtered, size=(size.height, size.width), antialias=False)
+
+        return out
+
+    def _preprocess(
+        self,
+        images: list["torch.Tensor"],
+        do_resize: bool,
+        size: SizeDict,
+        interpolation: Optional["F.InterpolationMode"],
+        do_pad: bool,
+        do_rescale: bool,
+        rescale_factor: float,
+        do_normalize: bool,
+        image_mean: Optional[Union[float, list[float]]],
+        image_std: Optional[Union[float, list[float]]],
+        disable_grouping: Optional[bool],
+        return_tensors: Optional[Union[str, TensorType]],
+        **kwargs,
+    ) -> BatchFeature:
+        # Group images by size for batched resizing
+        grouped_images, grouped_images_index = group_images_by_shape(images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+
+        for shape, stacked_images in grouped_images.items():
+            # Rescale images before other operations as done in original implementation
+            stacked_images = self.rescale_and_normalize(
+                stacked_images, do_rescale, rescale_factor, False, image_mean, image_std
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+
+        if do_pad:
+            processed_images = self.pad(processed_images, disable_grouping=disable_grouping)
+
+        grouped_images, grouped_images_index = group_images_by_shape(
+            processed_images, disable_grouping=disable_grouping
+        )
+        resized_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            if do_resize:
+                resized_stack = self.resize(
+                    image=stacked_images,
+                    size=size,
+                    interpolation=interpolation,
+                    input_data_format=ChannelDimension.FIRST,
+                )
+                resized_images_grouped[shape] = resized_stack
+        resized_images = reorder_images(resized_images_grouped, grouped_images_index)
+
+        # Group images by size for further processing
+        # Needed in case do_resize is False, or resize returns images with different sizes
+        grouped_images, grouped_images_index = group_images_by_shape(resized_images, disable_grouping=disable_grouping)
+        processed_images_grouped = {}
+        for shape, stacked_images in grouped_images.items():
+            # Fused rescale and normalize
+            stacked_images = self.rescale_and_normalize(
+                stacked_images, False, rescale_factor, do_normalize, image_mean, image_std
+            )
+            processed_images_grouped[shape] = stacked_images
+
+        processed_images = reorder_images(processed_images_grouped, grouped_images_index)
+
+        processed_images = torch.stack(processed_images, dim=0) if return_tensors else processed_images
+
+        return BatchFeature(data={"pixel_values": processed_images}, tensor_type=return_tensors)
+
+
+__all__ = ["Owlv2ImageProcessorFast"]

src/transformers/models/owlv2/processing_owlv2.py

@@ -56,19 +56,19 @@ class Owlv2ProcessorKwargs(ProcessingKwargs, total=False):
 
 class Owlv2Processor(ProcessorMixin):
     r"""
-    Constructs an Owlv2 processor which wraps [`Owlv2ImageProcessor`] and [`CLIPTokenizer`]/[`CLIPTokenizerFast`] into
+    Constructs an Owlv2 processor which wraps [`Owlv2ImageProcessor`]/[`Owlv2ImageProcessorFast`] and [`CLIPTokenizer`]/[`CLIPTokenizerFast`] into
     a single processor that inherits both the image processor and tokenizer functionalities. See the
     [`~OwlViTProcessor.__call__`] and [`~OwlViTProcessor.decode`] for more information.
 
     Args:
-        image_processor ([`Owlv2ImageProcessor`]):
+        image_processor ([`Owlv2ImageProcessor`, `Owlv2ImageProcessorFast`]):
             The image processor is a required input.
         tokenizer ([`CLIPTokenizer`, `CLIPTokenizerFast`]):
             The tokenizer is a required input.
     """
 
     attributes = ["image_processor", "tokenizer"]
-    image_processor_class = "Owlv2ImageProcessor"
+    image_processor_class = ("Owlv2ImageProcessor", "Owlv2ImageProcessorFast")
     tokenizer_class = ("CLIPTokenizer", "CLIPTokenizerFast")
 
     def __init__(self, image_processor, tokenizer, **kwargs):

tests/models/owlv2/test_image_processing_owlv2.py

@@ -16,7 +16,7 @@
 import unittest
 
 from transformers.testing_utils import require_torch, require_vision, slow
-from transformers.utils import is_torch_available, is_vision_available
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
 
 from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
 
@@ -29,6 +29,8 @@ if is_vision_available():
 if is_torch_available():
     import torch
 
+    from transformers import Owlv2ImageProcessorFast
+
 
 class Owlv2ImageProcessingTester:
     def __init__(
@@ -87,6 +89,7 @@ class Owlv2ImageProcessingTester:
 @require_vision
 class Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
     image_processing_class = Owlv2ImageProcessor if is_vision_available() else None
+    fast_image_processing_class = Owlv2ImageProcessorFast if is_torchvision_available() else None
 
     def setUp(self):
         super().setUp()
@@ -97,7 +100,8 @@ class Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
         return self.image_processor_tester.prepare_image_processor_dict()
 
     def test_image_processor_properties(self):
-        image_processing = self.image_processing_class(**self.image_processor_dict)
+        for image_processing_class in self.image_processor_list:
+            image_processing = image_processing_class(**self.image_processor_dict)
             self.assertTrue(hasattr(image_processing, "do_resize"))
             self.assertTrue(hasattr(image_processing, "size"))
             self.assertTrue(hasattr(image_processing, "do_normalize"))
@@ -105,17 +109,19 @@ class Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
             self.assertTrue(hasattr(image_processing, "image_std"))
 
     def test_image_processor_from_dict_with_kwargs(self):
-        image_processor = self.image_processing_class.from_dict(self.image_processor_dict)
+        for image_processing_class in self.image_processor_list:
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
             self.assertEqual(image_processor.size, {"height": 18, "width": 18})
 
-        image_processor = self.image_processing_class.from_dict(
+            image_processor = image_processing_class.from_dict(
                 self.image_processor_dict, size={"height": 42, "width": 42}
             )
             self.assertEqual(image_processor.size, {"height": 42, "width": 42})
 
     @slow
     def test_image_processor_integration_test(self):
-        processor = Owlv2ImageProcessor()
+        for image_processing_class in self.image_processor_list:
+            processor = image_processing_class()
 
             image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
             pixel_values = processor(image, return_tensors="pt").pixel_values
@@ -125,8 +131,9 @@ class Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
 
     @slow
     def test_image_processor_integration_test_resize(self):
+        for use_fast in [False, True]:
             checkpoint = "google/owlv2-base-patch16-ensemble"
-        processor = AutoProcessor.from_pretrained(checkpoint)
+            processor = AutoProcessor.from_pretrained(checkpoint, use_fast=use_fast)
             model = Owlv2ForObjectDetection.from_pretrained(checkpoint)
 
             image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
@@ -147,10 +154,7 @@ class Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
             results = processor.post_process_object_detection(outputs, threshold=0.2, target_sizes=[target_size])[0]
 
             boxes = results["boxes"]
-        self.assertTrue(
+            torch.testing.assert_close(boxes, expected_boxes, atol=1e-1, rtol=1e-1)
-            torch.allclose(boxes, expected_boxes, atol=1e-2),
-            f"Single image bounding boxes fail. Expected {expected_boxes}, got {boxes}",
-        )
 
             # batch of images
             inputs = processor(text=[text, text], images=[image, image], return_tensors="pt")
@@ -162,10 +166,7 @@ class Owlv2ImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
 
             for result in results:
                 boxes = result["boxes"]
-            self.assertTrue(
+                torch.testing.assert_close(boxes, expected_boxes, atol=1e-1, rtol=1e-1)
-                torch.allclose(boxes, expected_boxes, atol=1e-2),
-                f"Batch image bounding boxes fail. Expected {expected_boxes}, got {boxes}",
-            )
 
     @unittest.skip(reason="OWLv2 doesn't treat 4 channel PIL and numpy consistently yet")  # FIXME Amy
     def test_call_numpy_4_channels(self):