Add Image Processor Fast RT-DETR

docs/source/en/main_classes/image_processor.md

@@ -18,6 +18,49 @@ rendered properly in your Markdown viewer.
 
 An image processor is in charge of preparing input features for vision models and post processing their outputs. This includes transformations such as resizing, normalization, and conversion to PyTorch, TensorFlow, Flax and Numpy tensors. It may also include model specific post-processing such as converting logits to segmentation masks.
 
+Fast image processors are available for a few models and more will be added in the future. They are based on the [torchvision](https://pytorch.org/vision/stable/index.html) library and provide a significant speed-up, especially when processing on GPU.
+They have the same API as the base image processors and can be used as drop-in replacements.
+To use a fast image processor, you need to install the `torchvision` library, and set the `use_fast` argument to `True` when instantiating the image processor:
+
+```python
+from transformers import AutoImageProcessor
+
+processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50", use_fast=True)
+```
+
+When using a fast image processor, you can also set the `device` argument to specify the device on which the processing should be done. By default, the processing is done on the same device as the inputs if the inputs are tensors, or on the CPU otherwise.
+
+```python
+from torchvision.io import read_image
+from transformers import DetrImageProcessorFast
+
+images = read_image("image.jpg")
+processor = DetrImageProcessorFast.from_pretrained("facebook/detr-resnet-50")
+images_processed = processor(images, return_tensors="pt", device="cuda")
+```
+
+Here are some speed comparisons between the base and fast image processors for the `DETR` and `RT-DETR` models, and how they impact overall inference time:
+
+<div class="flex">
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/benchmark_results_full_pipeline_detr_fast_padded.png" />
+  </div>
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/benchmark_results_full_pipeline_detr_fast_batched_compiled.png" />
+  </div>
+</div>
+
+<div class="flex">
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/benchmark_results_full_pipeline_rt_detr_fast_single.png" />
+  </div>
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/benchmark_results_full_pipeline_rt_detr_fast_batched.png" />
+  </div>
+</div>
+
+These benchmarks were run on an [AWS EC2 g5.2xlarge instance](https://aws.amazon.com/ec2/instance-types/g5/), utilizing an NVIDIA A10G Tensor Core GPU.
+
 
 ## ImageProcessingMixin
 

docs/source/en/model_doc/rt_detr.md

@@ -46,7 +46,7 @@ Initially, an image is processed using a pre-trained convolutional neural networ
 >>> from PIL import Image
 >>> from transformers import RTDetrForObjectDetection, RTDetrImageProcessor
 
->>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg' 
+>>> url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
 >>> image = Image.open(requests.get(url, stream=True).raw)
 
 >>> image_processor = RTDetrImageProcessor.from_pretrained("PekingU/rtdetr_r50vd")
@@ -95,6 +95,12 @@ A list of official Hugging Face and community (indicated by 🌎) resources to h
     - preprocess
     - post_process_object_detection
 
+## RTDetrImageProcessorFast
+
+[[autodoc]] RTDetrImageProcessorFast
+    - preprocess
+    - post_process_object_detection
+
 ## RTDetrModel
 
 [[autodoc]] RTDetrModel

src/transformers/__init__.py

@@ -1228,7 +1228,7 @@
     _import_structure["models.poolformer"].extend(["PoolFormerFeatureExtractor", "PoolFormerImageProcessor"])
     _import_structure["models.pvt"].extend(["PvtImageProcessor"])
     _import_structure["models.qwen2_vl"].extend(["Qwen2VLImageProcessor"])
-    _import_structure["models.rt_detr"].extend(["RTDetrImageProcessor"])
+    _import_structure["models.rt_detr"].extend(["RTDetrImageProcessor", "RTDetrImageProcessorFast"])
     _import_structure["models.sam"].extend(["SamImageProcessor"])
     _import_structure["models.segformer"].extend(["SegformerFeatureExtractor", "SegformerImageProcessor"])
     _import_structure["models.seggpt"].extend(["SegGptImageProcessor"])
@@ -6152,7 +6152,7 @@
         )
         from .models.pvt import PvtImageProcessor
         from .models.qwen2_vl import Qwen2VLImageProcessor
-        from .models.rt_detr import RTDetrImageProcessor
+        from .models.rt_detr import RTDetrImageProcessor, RTDetrImageProcessorFast
         from .models.sam import SamImageProcessor
         from .models.segformer import SegformerFeatureExtractor, SegformerImageProcessor
         from .models.seggpt import SegGptImageProcessor

src/transformers/image_processing_utils_fast.py

@@ -15,14 +15,18 @@
 
 import functools
 from dataclasses import dataclass
+from typing import Any, Iterable, List, Optional, Tuple
 
 from .image_processing_utils import BaseImageProcessor
-from .utils.import_utils import is_torchvision_available
+from .utils.import_utils import is_torch_available, is_torchvision_available
 
 
 if is_torchvision_available():
     from torchvision.transforms import Compose
 
+if is_torch_available():
+    import torch
+
 
 @dataclass(frozen=True)
 class SizeDict:
@@ -66,3 +70,64 @@ def to_dict(self):
         encoder_dict = super().to_dict()
         encoder_dict.pop("_transform_params", None)
         return encoder_dict
+
+
+def get_image_size_for_max_height_width(
+    image_size: Tuple[int, int],
+    max_height: int,
+    max_width: int,
+) -> Tuple[int, int]:
+    """
+    Computes the output image size given the input image and the maximum allowed height and width. Keep aspect ratio.
+    Important, even if image_height < max_height and image_width < max_width, the image will be resized
+    to at least one of the edges be equal to max_height or max_width.
+
+    For example:
+        - input_size: (100, 200), max_height: 50, max_width: 50 -> output_size: (25, 50)
+        - input_size: (100, 200), max_height: 200, max_width: 500 -> output_size: (200, 400)
+
+    Args:
+        image_size (`Tuple[int, int]`):
+            The image to resize.
+        max_height (`int`):
+            The maximum allowed height.
+        max_width (`int`):
+            The maximum allowed width.
+    """
+    height, width = image_size
+    height_scale = max_height / height
+    width_scale = max_width / width
+    min_scale = min(height_scale, width_scale)
+    new_height = int(height * min_scale)
+    new_width = int(width * min_scale)
+    return new_height, new_width
+
+
+def safe_squeeze(tensor: "torch.Tensor", axis: Optional[int] = None) -> "torch.Tensor":
+    """
+    Squeezes a tensor, but only if the axis specified has dim 1.
+    """
+    if axis is None:
+        return tensor.squeeze()
+
+    try:
+        return tensor.squeeze(axis=axis)
+    except ValueError:
+        return tensor
+
+
+def max_across_indices(values: Iterable[Any]) -> List[Any]:
+    """
+    Return the maximum value across all indices of an iterable of values.
+    """
+    return [max(values_i) for values_i in zip(*values)]
+
+
+def get_max_height_width(images: List["torch.Tensor"]) -> Tuple[int]:
+    """
+    Get the maximum height and width across all images in a batch.
+    """
+
+    _, max_height, max_width = max_across_indices([img.shape for img in images])
+
+    return (max_height, max_width)

src/transformers/models/auto/image_processing_auto.py

@@ -123,7 +123,7 @@
             ("qwen2_vl", ("Qwen2VLImageProcessor",)),
             ("regnet", ("ConvNextImageProcessor",)),
             ("resnet", ("ConvNextImageProcessor",)),
-            ("rt_detr", "RTDetrImageProcessor"),
+            ("rt_detr", ("RTDetrImageProcessor", "RTDetrImageProcessorFast")),
             ("sam", ("SamImageProcessor",)),
             ("segformer", ("SegformerImageProcessor",)),
             ("seggpt", ("SegGptImageProcessor",)),