Add Fast Conditional-DETR Processor (#37071)

* Add Fast Conditional-DETR Processor

* Update image_processing_conditional_detr_fast.py

* Add modular_conditional_detr.py

* Update image_processing_conditional_detr_fast.py

* Update tests

* make fix

---------

Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>

docs/source/en/model_doc/conditional_detr.md

@@ -48,6 +48,11 @@ This model was contributed by [DepuMeng](https://huggingface.co/DepuMeng). The o
 
 [[autodoc]] ConditionalDetrImageProcessor
     - preprocess
+
+## ConditionalDetrImageProcessorFast
+
+[[autodoc]] ConditionalDetrImageProcessorFast
+    - preprocess
     - post_process_object_detection
     - post_process_instance_segmentation
     - post_process_semantic_segmentation

docs/source/ja/model_doc/conditional_detr.md

@@ -43,6 +43,11 @@ alt="描画" width="600"/>
 
 [[autodoc]] ConditionalDetrImageProcessor
     - preprocess
+
+## ConditionalDetrImageProcessorFast
+
+[[autodoc]] ConditionalDetrImageProcessorFast
+    - preprocess
     - post_process_object_detection
     - post_process_instance_segmentation
     - post_process_semantic_segmentation

src/transformers/models/auto/image_processing_auto.py

@@ -67,7 +67,7 @@ else:
             ("chinese_clip", ("ChineseCLIPImageProcessor", "ChineseCLIPImageProcessorFast")),
             ("clip", ("CLIPImageProcessor", "CLIPImageProcessorFast")),
             ("clipseg", ("ViTImageProcessor", "ViTImageProcessorFast")),
-            ("conditional_detr", ("ConditionalDetrImageProcessor",)),
+            ("conditional_detr", ("ConditionalDetrImageProcessor", "ConditionalDetrImageProcessorFast")),
             ("convnext", ("ConvNextImageProcessor", "ConvNextImageProcessorFast")),
             ("convnextv2", ("ConvNextImageProcessor", "ConvNextImageProcessorFast")),
             ("cvt", ("ConvNextImageProcessor", "ConvNextImageProcessorFast")),

src/transformers/models/conditional_detr/__init__.py

@@ -21,6 +21,7 @@ if TYPE_CHECKING:
     from .configuration_conditional_detr import *
     from .feature_extraction_conditional_detr import *
     from .image_processing_conditional_detr import *
+    from .image_processing_conditional_detr_fast import *
     from .modeling_conditional_detr import *
 else:
     import sys

src/transformers/models/conditional_detr/modular_conditional_detr.py

@@ -0,0 +1,137 @@
+from typing import List, Tuple, Union
+
+from transformers.models.detr.image_processing_detr_fast import DetrImageProcessorFast
+
+from ...image_transforms import (
+    center_to_corners_format,
+)
+from ...utils import (
+    TensorType,
+    is_torch_available,
+    logging,
+)
+
+
+if is_torch_available():
+    import torch
+
+
+logger = logging.get_logger(__name__)
+
+
+class ConditionalDetrImageProcessorFast(DetrImageProcessorFast):
+    def post_process(self, outputs, target_sizes):
+        """
+        Converts the output of [`ConditionalDetrForObjectDetection`] into the format expected by the Pascal VOC format (xmin, ymin, xmax, ymax).
+        Only supports PyTorch.
+
+        Args:
+            outputs ([`ConditionalDetrObjectDetectionOutput`]):
+                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.
+        """
+        logging.warning_once(
+            "`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.",
+        )
+
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+
+        if len(out_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")
+
+        prob = out_logits.sigmoid()
+        topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), 300, dim=1)
+        scores = topk_values
+        topk_boxes = torch.div(topk_indexes, out_logits.shape[2], rounding_mode="floor")
+        labels = topk_indexes % out_logits.shape[2]
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and 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)
+        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, threshold: float = 0.5, target_sizes: Union[TensorType, List[Tuple]] = None, top_k: int = 100
+    ):
+        """
+        Converts the raw output of [`ConditionalDetrForObjectDetection`] into final bounding boxes in (top_left_x,
+        top_left_y, bottom_right_x, bottom_right_y) format. Only supports PyTorch.
+
+        Args:
+            outputs ([`ConditionalDetrObjectDetectionOutput`]):
+                Raw outputs of the model.
+            threshold (`float`, *optional*):
+                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 left to None, predictions will not be resized.
+            top_k (`int`, *optional*, defaults to 100):
+                Keep only top k bounding boxes before filtering by thresholding.
+
+        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.
+        """
+        out_logits, out_bbox = outputs.logits, outputs.pred_boxes
+
+        if target_sizes is not None:
+            if len(out_logits) != len(target_sizes):
+                raise ValueError(
+                    "Make sure that you pass in as many target sizes as the batch dimension of the logits"
+                )
+
+        prob = out_logits.sigmoid()
+        prob = prob.view(out_logits.shape[0], -1)
+        k_value = min(top_k, prob.size(1))
+        topk_values, topk_indexes = torch.topk(prob, k_value, dim=1)
+        scores = topk_values
+        topk_boxes = torch.div(topk_indexes, out_logits.shape[2], rounding_mode="floor")
+        labels = topk_indexes % out_logits.shape[2]
+        boxes = center_to_corners_format(out_bbox)
+        boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
+
+        # and from relative [0, 1] to absolute [0, height] coordinates
+        if target_sizes is not None:
+            if isinstance(target_sizes, List):
+                img_h = torch.Tensor([i[0] for i in target_sizes])
+                img_w = torch.Tensor([i[1] for i in target_sizes])
+            else:
+                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 = []
+        for s, l, b in zip(scores, labels, boxes):
+            score = s[s > threshold]
+            label = l[s > threshold]
+            box = b[s > threshold]
+            results.append({"scores": score, "labels": label, "boxes": box})
+
+        return results
+
+    def post_process_segmentation():
+        raise NotImplementedError("Segmentation post-processing is not implemented for Conditional DETR yet.")
+
+    def post_process_instance():
+        raise NotImplementedError("Instance post-processing is not implemented for Conditional DETR yet.")
+
+    def post_process_panoptic():
+        raise NotImplementedError("Panoptic post-processing is not implemented for Conditional DETR yet.")
+
+
+__all__ = ["ConditionalDetrImageProcessorFast"]

tests/models/conditional_detr/test_image_processing_conditional_detr.py

@@ -20,7 +20,7 @@ import unittest
 import numpy as np
 
 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 AnnotationFormatTestMixin, ImageProcessingTestMixin, prepare_image_inputs
 
@@ -33,6 +33,9 @@ if is_vision_available():
 
     from transformers import ConditionalDetrImageProcessor
 
+    if is_torchvision_available():
+        from transformers import ConditionalDetrImageProcessorFast
+
 
 class ConditionalDetrImageProcessingTester:
     def __init__(
@@ -132,6 +135,7 @@ class ConditionalDetrImageProcessingTester:
 @require_vision
 class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcessingTestMixin, unittest.TestCase):
     image_processing_class = ConditionalDetrImageProcessor if is_vision_available() else None
+    fast_image_processing_class = ConditionalDetrImageProcessorFast if is_torchvision_available() else None
 
     def setUp(self):
         super().setUp()
@@ -142,7 +146,8 @@ class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcess
         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, "image_mean"))
             self.assertTrue(hasattr(image_processing, "image_std"))
             self.assertTrue(hasattr(image_processing, "do_normalize"))
@@ -150,11 +155,12 @@ class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcess
             self.assertTrue(hasattr(image_processing, "size"))
 
     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, {"shortest_edge": 18, "longest_edge": 1333})
             self.assertEqual(image_processor.do_pad, True)
 
-        image_processor = self.image_processing_class.from_dict(
+            image_processor = image_processing_class.from_dict(
                 self.image_processor_dict, size=42, max_size=84, pad_and_return_pixel_mask=False
             )
             self.assertEqual(image_processor.size, {"shortest_edge": 42, "longest_edge": 84})
@@ -169,8 +175,9 @@ class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcess
 
         target = {"image_id": 39769, "annotations": target}
 
+        for image_processing_class in self.image_processor_list:
             # encode them
-        image_processing = ConditionalDetrImageProcessor.from_pretrained("microsoft/conditional-detr-resnet-50")
+            image_processing = image_processing_class.from_pretrained("microsoft/conditional-detr-resnet-50")
             encoding = image_processing(images=image, annotations=target, return_tensors="pt")
 
             # verify pixel values
@@ -215,8 +222,9 @@ class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcess
 
         masks_path = pathlib.Path("./tests/fixtures/tests_samples/COCO/coco_panoptic")
 
+        for image_processing_class in self.image_processor_list:
             # encode them
-        image_processing = ConditionalDetrImageProcessor(format="coco_panoptic")
+            image_processing = image_processing_class(format="coco_panoptic")
             encoding = image_processing(images=image, annotations=target, masks_path=masks_path, return_tensors="pt")
 
             # verify pixel values
@@ -245,7 +253,8 @@ class ConditionalDetrImageProcessingTest(AnnotationFormatTestMixin, ImageProcess
             torch.testing.assert_close(encoding["labels"][0]["class_labels"], expected_class_labels)
             # verify masks
             expected_masks_sum = 822873
-        self.assertEqual(encoding["labels"][0]["masks"].sum().item(), expected_masks_sum)
+            relative_error = torch.abs(encoding["labels"][0]["masks"].sum() - expected_masks_sum) / expected_masks_sum
+            self.assertTrue(relative_error < 1e-3)
             # verify orig_size
             expected_orig_size = torch.tensor([480, 640])
             torch.testing.assert_close(encoding["labels"][0]["orig_size"], expected_orig_size)