New model support RTDETR (#29077)

* fill out docs string in configuration
https://github.com/huggingface/transformers/pull/29077/files/75dcd3a0e82cca36f12178b65bbd071ab7b25088#r1506391856



* reduce the input image size for the tests

* remove the unappropriate tests

* only 5 failes exists

* make style

* fill up missed architecture for object detection in docs

* fix auto modeling

* simple fix in missing import

* major change including backbone refactor and objectdetectionoutput refactor

* minor fix only 4 fails left

* intermediate fix

* revert __init__.py

* revert __init__.py

* make style

* fixes in pr_docs

* intermediate fix

* make style

* two fixes

* pass doctest

* only one fix left

* intermediate commit

* all fixed

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/convert_rt_detr_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update tests/models/rt_detr/test_modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* function class above the model definition in dice_loss

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* simple fix

* layernorm add config.layer_norm_eps

* fix inputs_docstring

* make style

* simple fix

* add custom coco loading test in image_processor

* fix error in BaseModelOutput
https://github.com/huggingface/transformers/pull/29077#discussion_r1516657790



* simple typo

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* intermediate fix

* fix with load_backbone format

* remove unused configuration

* 3 fix test left

* make style

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: Sounak Dey <dey.sounak@gmail.com>

* change last_hidden_state to first index

* all pass fix
TO DO: minor update in comments

* make fix-copies

* remove deepcopy

* pr_document fix

* revert deepcopy due to the issue of unexpceted behavior in decoderlayer

* add atol in final

* add no_split_module

* _no_split_modules = None

* device transfer for model parallelism

* minor fix

* make fix-copies

* fix typo

* add test_image_processor with post_processing

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* add config in RTDETRPredictionHead

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set lru_cache with max_size 32

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* add lru_cache import and configuration change

* change the order of definition

* make fix-copies

* add docs and change config error

* revert strange make-fix

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* test pass

* fix get_clones related and remove deepcopy

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* nit for paper section

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* rename denoising related parameters

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* check the image transformation logic

* make style

* make style

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* pe_encoding -> positional_encoding_temperature

* remove TODO

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* remove eval_idx since transformer DETR is giving all decoder output

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* change variable name

* make style and docs import update

* Revert "Update src/transformers/models/rt_detr/image_processing_rt_detr.py"

This reverts commit 74aa3e1de0ca0cd3d354161d38ef28b4389c0eee.

* fix typo

* add postprocessing in docs

* move import scipy to top

* change varaible name

* make fix-copies

* remove eval_idx in test

* move to after first sentence

* update image_processor since box loss requires normalized one

* change appropriate name to auxiliary_outputs

* Update src/transformers/models/rt_detr/__init__.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/rt_detr/__init__.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/rt_detr.md
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/rt_detr.md
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* make style

* remove panoptic related comments

* make style

* revert valid_processor_keys

* fix aux related test

* make style

* change origination from config to backbone API

* enable the dn_loss

* fix test and conversion

* renewal weight initialization

* change initializer_range

* make fix-up

* fix the loss issue in the auxiliary output and denoising part

* change weight loss to original RTDETR

* fix in initialization

* sync shape format of dn and aux

* make style

* stable fine-tuning and compatible conversion for resnet101

* make style

* skip input_embed

* change encoder related variable

* enable converting rtdetr_r101

* add r101 related conversion code

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update docs/source/en/model_doc/rt_detr.md
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/image_processing_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* change name _shape to _reshape

* Update src/transformers/__init__.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* maket style

* make fix-copies

* remove deprecated import

* more fix

* remove last_hidden_state for task-specific model

* Revert "remove last_hidden_state for task-specific model"

This reverts commit ccb7a34051d69b9fc7aa17ed8644664d3fdbdaca.

* minore change in convert

* remove print

* make style and fix-copies

* add custom rtdetr backbone for r18, r34

* remove print

* change copied

* add pad_size

* make style

* change layertype to optional to pass the CI

* make style

* add test in modeling_resnet_rt_detr

* make fix-copies

* skip tmp file test

* fix comment

* add docs

* change to modeling_resnet file format

* enabling resnet50 above

* Update src/transformers/models/rt_detr/modeling_rt_detr.py
Co-authored-by: Jason Wu <jasonkit@users.noreply.github.com>

* enable all the rtdetr model :)

* finish except CI

* add RTDetrResNetBackbone

* make fix-copies

* fix
TO DO: CI enable

* make style

* rename test

* add docs

* add special fix

* revert resnet

* Update src/transformers/models/rt_detr/modeling_rt_detr_resnet.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* add more comment

* remove swin comment

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* rename convert and add verify backbone

* Update docs/source/en/_toctree.yml
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/rt_detr.md
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/rt_detr.md
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* make style

* requests for docs

* more general test docs

* general script docs

* make fix-copies

* final commit

* Revert "Update src/transformers/models/rt_detr/configuration_rt_detr.py"

This reverts commit d136225cd3f64f510d303ce1d227698174f43fff.

* skip test_model_get_set_embeddings

* remove target

* add changes

* make fix-copies

* remove decoder_attention_mask

* add load_backbone function for auto_backbone

* remove comment

* fix repo name

* Update src/transformers/models/rt_detr/configuration_rt_detr.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* final commit

* remove unused downsample_in_bottleneck

* new test for autobackbone

* change to appropriate indices

* test fix

* fix dict in test_image_processor

* fix test

* [run-slow] rt_detr, rt_detr_resnet

* change the slow test

* [run-slow] rt_detr

* [run-slow] rt_detr, rt_detr_resnet

* make in to same cuda in CSPRepLayer

* [run-slow] rt_detr, rt_detr_resnet

---------
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: Sounak Dey <dey.sounak@gmail.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Jason Wu <jasonkit@users.noreply.github.com>
Co-authored-by: ChoiSangBum <choisangbum@ChoiSangBumui-MacBookPro.local>

docs/source/en/_toctree.yml

@@ -627,6 +627,8 @@
         title: RegNet
       - local: model_doc/resnet
         title: ResNet
+      - local: model_doc/rt_detr
+        title: RT-DETR
       - local: model_doc/segformer
         title: SegFormer
       - local: model_doc/seggpt

docs/source/en/index.md

@@ -262,6 +262,8 @@ Flax), PyTorch, and/or TensorFlow.
 |          [RoBERTa-PreLayerNorm](model_doc/roberta-prelayernorm)          |       ✅        |         ✅         |      ✅      |
 |                      [RoCBert](model_doc/roc_bert)                       |       ✅        |         ❌         |      ❌      |
 |                      [RoFormer](model_doc/roformer)                      |       ✅        |         ✅         |      ✅      |
+|                       [RT-DETR](model_doc/rt_detr)                       |       ✅        |         ❌         |      ❌      |
+|                [RT-DETR-ResNet](model_doc/rt_detr_resnet)                |       ✅        |         ❌         |      ❌      |
 |                          [RWKV](model_doc/rwkv)                          |       ✅        |         ❌         |      ❌      |
 |                           [SAM](model_doc/sam)                           |       ✅        |         ✅         |      ❌      |
 |                  [SeamlessM4T](model_doc/seamless_m4t)                   |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/rt_detr.md

+<!--Copyright 2024 The HuggingFace 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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# RT-DETR
+
+## Overview
+
+
+The RT-DETR model was proposed in [DETRs Beat YOLOs on Real-time Object Detection](https://arxiv.org/abs/2304.08069) by Wenyu Lv, Yian Zhao, Shangliang Xu, Jinman Wei, Guanzhong Wang, Cheng Cui, Yuning Du, Qingqing Dang, Yi Liu.
+
+RT-DETR is an object detection model that stands for "Real-Time DEtection Transformer." This model is designed to perform object detection tasks with a focus on achieving real-time performance while maintaining high accuracy. Leveraging the transformer architecture, which has gained significant popularity in various fields of deep learning, RT-DETR processes images to identify and locate multiple objects within them.
+
+The abstract from the paper is the following:
+
+*Recently, end-to-end transformer-based detectors (DETRs) have achieved remarkable performance. However, the issue of the high computational cost of DETRs has not been effectively addressed, limiting their practical application and preventing them from fully exploiting the benefits of no post-processing, such as non-maximum suppression (NMS). In this paper, we first analyze the influence of NMS in modern real-time object detectors on inference speed, and establish an end-to-end speed benchmark. To avoid the inference delay caused by NMS, we propose a Real-Time DEtection TRansformer (RT-DETR), the first real-time end-to-end object detector to our best knowledge. Specifically, we design an efficient hybrid encoder to efficiently process multi-scale features by decoupling the intra-scale interaction and cross-scale fusion, and propose IoU-aware query selection to improve the initialization of object queries. In addition, our proposed detector supports flexibly adjustment of the inference speed by using different decoder layers without the need for retraining, which facilitates the practical application of real-time object detectors. Our RT-DETR-L achieves 53.0% AP on COCO val2017 and 114 FPS on T4 GPU, while RT-DETR-X achieves 54.8% AP and 74 FPS, outperforming all YOLO detectors of the same scale in both speed and accuracy. Furthermore, our RT-DETR-R50 achieves 53.1% AP and 108 FPS, outperforming DINO-Deformable-DETR-R50 by 2.2% AP in accuracy and by about 21 times in FPS.*
+
+The model version was contributed by [rafaelpadilla](https://huggingface.co/rafaelpadilla) and [sangbumchoi](https://github.com/SangbumChoi). The original code can be found [here](https://github.com/lyuwenyu/RT-DETR/).
+
+
+## Usage tips
+
+Initially, an image is processed using a pre-trained convolutional neural network, specifically a Resnet-D variant as referenced in the original code. This network extracts features from the final three layers of the architecture. Following this, a hybrid encoder is employed to convert the multi-scale features into a sequential array of image features. Then, a decoder, equipped with auxiliary prediction heads is used to refine the object queries. This process facilitates the direct generation of bounding boxes, eliminating the need for any additional post-processing to acquire the logits and coordinates for the bounding boxes.
+
+```py
+from transformers import RTDetrForObjectDetection, RTDetrImageProcessor
+from PIL import Image
+import json
+import torch
+import requests
+
+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")
+model = RTDetrForObjectDetection.from_pretrained("PekingU/rtdetr_r50vd")
+
+inputs = image_processor(images=image, return_tensors="pt")
+
+with torch.no_grad():
+   outputs = model(**inputs)
+
+results = image_processor.post_process_object_detection(outputs, target_sizes=torch.tensor([image.size[::-1]), threshold=0.3)
+```
+
+## RTDetrConfig
+
+[[autodoc]] RTDetrConfig
+
+## RTDetrResNetConfig
+
+[[autodoc]] RTDetrResNetConfig
+
+## RTDetrImageProcessor
+
+[[autodoc]] RTDetrImageProcessor
+    - preprocess
+    - post_process_object_detection
+
+## RTDetrModel
+
+[[autodoc]] RTDetrModel
+    - forward
+
+## RTDetrForObjectDetection
+
+[[autodoc]] RTDetrForObjectDetection
+    - forward
+
+## RTDetrResNetBackbone
+
+[[autodoc]] RTDetrResNetBackbone
+    - forward
\ No newline at end of file

src/transformers/__init__.py

@@ -654,6 +654,7 @@ _import_structure = {
         "RoFormerConfig",
         "RoFormerTokenizer",
     ],
+    "models.rt_detr": ["RTDetrConfig", "RTDetrResNetConfig"],
     "models.rwkv": ["RwkvConfig"],
     "models.sam": [
         "SamConfig",
@@ -1153,6 +1154,7 @@ else:
     _import_structure["models.pix2struct"].extend(["Pix2StructImageProcessor"])
     _import_structure["models.poolformer"].extend(["PoolFormerFeatureExtractor", "PoolFormerImageProcessor"])
     _import_structure["models.pvt"].extend(["PvtImageProcessor"])
+    _import_structure["models.rt_detr"].extend(["RTDetrImageProcessor"])
     _import_structure["models.sam"].extend(["SamImageProcessor"])
     _import_structure["models.segformer"].extend(["SegformerFeatureExtractor", "SegformerImageProcessor"])
     _import_structure["models.seggpt"].extend(["SegGptImageProcessor"])
@@ -3004,6 +3006,15 @@ else:
             "load_tf_weights_in_roformer",
         ]
     )
+    _import_structure["models.rt_detr"].extend(
+        [
+            "RTDetrForObjectDetection",
+            "RTDetrModel",
+            "RTDetrPreTrainedModel",
+            "RTDetrResNetBackbone",
+            "RTDetrResNetPreTrainedModel",
+        ]
+    )
     _import_structure["models.rwkv"].extend(
         [
             "RwkvForCausalLM",
@@ -5270,6 +5281,10 @@ if TYPE_CHECKING:
         RoFormerConfig,
         RoFormerTokenizer,
     )
+    from .models.rt_detr import (
+        RTDetrConfig,
+        RTDetrResNetConfig,
+    )
     from .models.rwkv import RwkvConfig
     from .models.sam import (
         SamConfig,
@@ -5792,6 +5807,7 @@ if TYPE_CHECKING:
             PoolFormerImageProcessor,
         )
         from .models.pvt import PvtImageProcessor
+        from .models.rt_detr import RTDetrImageProcessor
         from .models.sam import SamImageProcessor
         from .models.segformer import SegformerFeatureExtractor, SegformerImageProcessor
         from .models.seggpt import SegGptImageProcessor
@@ -7295,6 +7311,13 @@ if TYPE_CHECKING:
             RoFormerPreTrainedModel,
             load_tf_weights_in_roformer,
         )
+        from .models.rt_detr import (
+            RTDetrForObjectDetection,
+            RTDetrModel,
+            RTDetrPreTrainedModel,
+            RTDetrResNetBackbone,
+            RTDetrResNetPreTrainedModel,
+        )
         from .models.rwkv import (
             RwkvForCausalLM,
             RwkvModel,

src/transformers/models/__init__.py

@@ -193,6 +193,7 @@ from . import (
     roberta_prelayernorm,
     roc_bert,
     roformer,
+    rt_detr,
     rwkv,
     sam,
     seamless_m4t,

src/transformers/models/auto/configuration_auto.py

@@ -214,6 +214,8 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("roberta-prelayernorm", "RobertaPreLayerNormConfig"),
         ("roc_bert", "RoCBertConfig"),
         ("roformer", "RoFormerConfig"),
+        ("rt_detr", "RTDetrConfig"),
+        ("rt_detr_resnet", "RTDetrResNetConfig"),
         ("rwkv", "RwkvConfig"),
         ("sam", "SamConfig"),
         ("seamless_m4t", "SeamlessM4TConfig"),
@@ -499,6 +501,8 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("roberta-prelayernorm", "RoBERTa-PreLayerNorm"),
         ("roc_bert", "RoCBert"),
         ("roformer", "RoFormer"),
+        ("rt_detr", "RT-DETR"),
+        ("rt_detr_resnet", "RT-DETR-ResNet"),
         ("rwkv", "RWKV"),
         ("sam", "SAM"),
         ("seamless_m4t", "SeamlessM4T"),
@@ -623,6 +627,7 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict(
         ("clip_vision_model", "clip"),
         ("siglip_vision_model", "siglip"),
         ("chinese_clip_vision_model", "chinese_clip"),
+        ("rt_detr_resnet", "rt_detr"),
     ]
 )
 

src/transformers/models/auto/image_processing_auto.py

@@ -114,6 +114,7 @@ else:
             ("pvt_v2", ("PvtImageProcessor",)),
             ("regnet", ("ConvNextImageProcessor",)),
             ("resnet", ("ConvNextImageProcessor",)),
+            ("rt_detr", "RTDetrImageProcessor"),
             ("sam", ("SamImageProcessor",)),
             ("segformer", ("SegformerImageProcessor",)),
             ("seggpt", ("SegGptImageProcessor",)),

src/transformers/models/auto/modeling_auto.py

@@ -202,6 +202,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("roberta-prelayernorm", "RobertaPreLayerNormModel"),
         ("roc_bert", "RoCBertModel"),
         ("roformer", "RoFormerModel"),
+        ("rt_detr", "RTDetrModel"),
         ("rwkv", "RwkvModel"),
         ("sam", "SamModel"),
         ("seamless_m4t", "SeamlessM4TModel"),
@@ -765,6 +766,7 @@ MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
         ("deformable_detr", "DeformableDetrForObjectDetection"),
         ("deta", "DetaForObjectDetection"),
         ("detr", "DetrForObjectDetection"),
+        ("rt_detr", "RTDetrForObjectDetection"),
         ("table-transformer", "TableTransformerForObjectDetection"),
         ("yolos", "YolosForObjectDetection"),
     ]
@@ -1252,6 +1254,7 @@ MODEL_FOR_BACKBONE_MAPPING_NAMES = OrderedDict(
         ("nat", "NatBackbone"),
         ("pvt_v2", "PvtV2Backbone"),
         ("resnet", "ResNetBackbone"),
+        ("rt_detr_resnet", "RTDetrResNetBackbone"),
         ("swin", "SwinBackbone"),
         ("swinv2", "Swinv2Backbone"),
         ("timm_backbone", "TimmBackbone"),

src/transformers/models/deformable_detr/modeling_deformable_detr.py

@@ -29,22 +29,24 @@ from torch.autograd import Function
 from torch.autograd.function import once_differentiable
 
 from ...activations import ACT2FN
-from ...file_utils import (
+from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
+from ...modeling_outputs import BaseModelOutput
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import meshgrid
+from ...utils import (
     ModelOutput,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
+    is_accelerate_available,
+    is_ninja_available,
     is_scipy_available,
     is_timm_available,
     is_torch_cuda_available,
     is_vision_available,
+    logging,
     replace_return_docstrings,
     requires_backends,
 )
-from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
-from ...modeling_outputs import BaseModelOutput
-from ...modeling_utils import PreTrainedModel
-from ...pytorch_utils import meshgrid
-from ...utils import is_accelerate_available, is_ninja_available, logging
 from ...utils.backbone_utils import load_backbone
 from .configuration_deformable_detr import DeformableDetrConfig
 

src/transformers/models/rt_detr/__init__.py

+# Copyright 2024 The HuggingFace 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.
+
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+
+
+_import_structure = {"configuration_rt_detr": ["RTDetrConfig"], "configuration_rt_detr_resnet": ["RTDetrResNetConfig"]}
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_rt_detr"] = ["RTDetrImageProcessor"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_rt_detr"] = [
+        "RTDetrForObjectDetection",
+        "RTDetrModel",
+        "RTDetrPreTrainedModel",
+    ]
+    _import_structure["modeling_rt_detr_resnet"] = [
+        "RTDetrResNetBackbone",
+        "RTDetrResNetPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_rt_detr import RTDetrConfig
+    from .configuration_rt_detr_resnet import RTDetrResNetConfig
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_rt_detr import RTDetrImageProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_rt_detr import (
+            RTDetrForObjectDetection,
+            RTDetrModel,
+            RTDetrPreTrainedModel,
+        )
+        from .modeling_rt_detr_resnet import (
+            RTDetrResNetBackbone,
+            RTDetrResNetPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/rt_detr/configuration_rt_detr.py

+# coding=utf-8
+# Copyright 2024 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.
+"""RT-DETR model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ...utils.backbone_utils import verify_backbone_config_arguments
+from ..auto import CONFIG_MAPPING
+from .configuration_rt_detr_resnet import RTDetrResNetConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+class RTDetrConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`RTDetrModel`]. It is used to instantiate a
+    RT-DETR model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the RT-DETR
+    [checkpoing/todo](https://huggingface.co/checkpoing/todo) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        initializer_range (`float`, *optional*, defaults to 0.01):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the layer normalization layers.
+        batch_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the batch normalization layers.
+        backbone_config (`Dict`, *optional*, defaults to `RTDetrResNetConfig()`):
+            The configuration of the backbone model.
+        backbone (`str`, *optional*):
+            Name of backbone to use when `backbone_config` is `None`. If `use_pretrained_backbone` is `True`, this
+            will load the corresponding pretrained weights from the timm or transformers library. If `use_pretrained_backbone`
+            is `False`, this loads the backbone's config and uses that to initialize the backbone with random weights.
+        use_pretrained_backbone (`bool`, *optional*, defaults to `False`):
+            Whether to use pretrained weights for the backbone.
+        use_timm_backbone (`bool`, *optional*, defaults to `False`):
+            Whether to load `backbone` from the timm library. If `False`, the backbone is loaded from the transformers
+            library.
+        backbone_kwargs (`dict`, *optional*):
+            Keyword arguments to be passed to AutoBackbone when loading from a checkpoint
+            e.g. `{'out_indices': (0, 1, 2, 3)}`. Cannot be specified if `backbone_config` is set.
+        encoder_hidden_dim (`int`, *optional*, defaults to 256):
+            Dimension of the layers in hybrid encoder.
+        encoder_in_channels (`list`, *optional*, defaults to `[512, 1024, 2048]`):
+            Multi level features input for encoder.
+        feat_strides (`List[int]`, *optional*, defaults to `[8, 16, 32]`):
+            Strides used in each feature map.
+        encoder_layers (`int`, *optional*, defaults to 1):
+            Total of layers to be used by the encoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The ratio for all dropout layers.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        encode_proj_layers (`List[int]`, *optional*, defaults to `[2]`):
+            Indexes of the projected layers to be used in the encoder.
+        positional_encoding_temperature (`int`, *optional*, defaults to 10000):
+            The temperature parameter used to create the positional encodings.
+        encoder_activation_function (`str`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        activation_function (`str`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the general layer. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        eval_size (`Tuple[int, int]`, *optional*):
+            Height and width used to computes the effective height and width of the position embeddings after taking
+            into account the stride.
+        normalize_before (`bool`, *optional*, defaults to `False`):
+            Determine whether to apply layer normalization in the transformer encoder layer before self-attention and
+            feed-forward modules.
+        hidden_expansion (`float`, *optional*, defaults to 1.0):
+            Expansion ratio to enlarge the dimension size of RepVGGBlock and CSPRepLayer.
+        d_model (`int`, *optional*, defaults to 256):
+            Dimension of the layers exclude hybrid encoder.
+        num_queries (`int`, *optional*, defaults to 300):
+            Number of object queries.
+        decoder_in_channels (`list`, *optional*, defaults to `[256, 256, 256]`):
+            Multi level features dimension for decoder
+        decoder_ffn_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        num_feature_levels (`int`, *optional*, defaults to 3):
+            The number of input feature levels.
+        decoder_n_points (`int`, *optional*, defaults to 4):
+            The number of sampled keys in each feature level for each attention head in the decoder.
+        decoder_layers (`int`, *optional*, defaults to 6):
+            Number of decoder layers.
+        decoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_activation_function (`str`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the decoder. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        num_denoising (`int`, *optional*, defaults to 100):
+            The total number of denoising tasks or queries to be used for contrastive denoising.
+        label_noise_ratio (`float`, *optional*, defaults to 0.5):
+            The fraction of denoising labels to which random noise should be added.
+        box_noise_scale (`float`, *optional*, defaults to 1.0):
+            Scale or magnitude of noise to be added to the bounding boxes.
+        learn_initial_query (`bool`, *optional*, defaults to `False`):
+            Indicates whether the initial query embeddings for the decoder should be learned during training
+        anchor_image_size (`Tuple[int, int]`, *optional*, defaults to `[640, 640]`):
+            Height and width of the input image used during evaluation to generate the bounding box anchors.
+        disable_custom_kernels (`bool`, *optional*, defaults to `True`):
+            Whether to disable custom kernels.
+        with_box_refine (`bool`, *optional*, defaults to `True`):
+            Whether to apply iterative bounding box refinement, where each decoder layer refines the bounding boxes
+            based on the predictions from the previous layer.
+        is_encoder_decoder (`bool`, *optional*, defaults to `True`):
+            Whether the architecture has an encoder decoder structure.
+        matcher_alpha (`float`, *optional*, defaults to 0.25):
+            Parameter alpha used by the Hungarian Matcher.
+        matcher_gamma (`float`, *optional*, defaults to 2.0):
+            Parameter gamma used by the Hungarian Matcher.
+        matcher_class_cost (`float`, *optional*, defaults to 2.0):
+            The relative weight of the class loss used by the Hungarian Matcher.
+        matcher_bbox_cost (`float`, *optional*, defaults to 5.0):
+            The relative weight of the bounding box loss used by the Hungarian Matcher.
+        matcher_giou_cost (`float`, *optional*, defaults to 2.0):
+            The relative weight of the giou loss of used by the Hungarian Matcher.
+        use_focal_loss (`bool`, *optional*, defaults to `True`):
+            Parameter informing if focal focal should be used.
+        auxiliary_loss (`bool`, *optional*, defaults to `True`):
+            Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
+        focal_loss_alpha (`float`, *optional*, defaults to 0.75):
+            Parameter alpha used to compute the focal loss.
+        focal_loss_gamma (`float`, *optional*, defaults to 2.0):
+            Parameter gamma used to compute the focal loss.
+        weight_loss_vfl (`float`, *optional*, defaults to 1.0):
+            Relative weight of the varifocal loss in the object detection loss.
+        weight_loss_bbox (`float`, *optional*, defaults to 5.0):
+            Relative weight of the L1 bounding box loss in the object detection loss.
+        weight_loss_giou (`float`, *optional*, defaults to 2.0):
+            Relative weight of the generalized IoU loss in the object detection loss.
+        eos_coefficient (`float`, *optional*, defaults to 0.0001):
+            Relative classification weight of the 'no-object' class in the object detection loss.
+
+    Examples:
+
+    ```python
+    >>> from transformers import RTDetrConfig, RTDetrModel
+
+    >>> # Initializing a RT-DETR configuration
+    >>> configuration = RTDetrConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = RTDetrModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "rt_detr"
+    layer_types = ["basic", "bottleneck"]
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    def __init__(
+        self,
+        initializer_range=0.01,
+        layer_norm_eps=1e-5,
+        batch_norm_eps=1e-5,
+        # backbone
+        backbone_config=None,
+        backbone=None,
+        use_pretrained_backbone=False,
+        use_timm_backbone=False,
+        backbone_kwargs=None,
+        # encoder HybridEncoder
+        encoder_hidden_dim=256,
+        encoder_in_channels=[512, 1024, 2048],
+        feat_strides=[8, 16, 32],
+        encoder_layers=1,
+        encoder_ffn_dim=1024,
+        encoder_attention_heads=8,
+        dropout=0.0,
+        activation_dropout=0.0,
+        encode_proj_layers=[2],
+        positional_encoding_temperature=10000,
+        encoder_activation_function="gelu",
+        activation_function="silu",
+        eval_size=None,
+        normalize_before=False,
+        hidden_expansion=1.0,
+        # decoder RTDetrTransformer
+        d_model=256,
+        num_queries=300,
+        decoder_in_channels=[256, 256, 256],
+        decoder_ffn_dim=1024,
+        num_feature_levels=3,
+        decoder_n_points=4,
+        decoder_layers=6,
+        decoder_attention_heads=8,
+        decoder_activation_function="relu",
+        attention_dropout=0.0,
+        num_denoising=100,
+        label_noise_ratio=0.5,
+        box_noise_scale=1.0,
+        learn_initial_query=False,
+        anchor_image_size=[640, 640],
+        disable_custom_kernels=True,
+        with_box_refine=True,
+        is_encoder_decoder=True,
+        # Loss
+        matcher_alpha=0.25,
+        matcher_gamma=2.0,
+        matcher_class_cost=2.0,
+        matcher_bbox_cost=5.0,
+        matcher_giou_cost=2.0,
+        use_focal_loss=True,
+        auxiliary_loss=True,
+        focal_loss_alpha=0.75,
+        focal_loss_gamma=2.0,
+        weight_loss_vfl=1.0,
+        weight_loss_bbox=5.0,
+        weight_loss_giou=2.0,
+        eos_coefficient=1e-4,
+        **kwargs,
+    ):
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.batch_norm_eps = batch_norm_eps
+        # backbone
+        if backbone_config is None and backbone is None:
+            logger.info(
+                "`backbone_config` and `backbone` are `None`. Initializing the config with the default `RTDetr-ResNet` backbone."
+            )
+            backbone_config = RTDetrResNetConfig(
+                num_channels=3,
+                embedding_size=64,
+                hidden_sizes=[256, 512, 1024, 2048],
+                depths=[3, 4, 6, 3],
+                layer_type="bottleneck",
+                hidden_act="relu",
+                downsample_in_first_stage=False,
+                downsample_in_bottleneck=False,
+                out_features=None,
+                out_indices=[2, 3, 4],
+            )
+        elif isinstance(backbone_config, dict):
+            backbone_model_type = backbone_config.pop("model_type")
+            config_class = CONFIG_MAPPING[backbone_model_type]
+            backbone_config = config_class.from_dict(backbone_config)
+
+        verify_backbone_config_arguments(
+            use_timm_backbone=use_timm_backbone,
+            use_pretrained_backbone=use_pretrained_backbone,
+            backbone=backbone,
+            backbone_config=backbone_config,
+            backbone_kwargs=backbone_kwargs,
+        )
+
+        self.backbone_config = backbone_config
+        self.backbone = backbone
+        self.use_pretrained_backbone = use_pretrained_backbone
+        self.use_timm_backbone = use_timm_backbone
+        self.backbone_kwargs = backbone_kwargs
+        # encoder
+        self.encoder_hidden_dim = encoder_hidden_dim
+        self.encoder_in_channels = encoder_in_channels
+        self.feat_strides = feat_strides
+        self.encoder_attention_heads = encoder_attention_heads
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.dropout = dropout
+        self.activation_dropout = activation_dropout
+        self.encode_proj_layers = encode_proj_layers
+        self.encoder_layers = encoder_layers
+        self.positional_encoding_temperature = positional_encoding_temperature
+        self.eval_size = eval_size
+        self.normalize_before = normalize_before
+        self.encoder_activation_function = encoder_activation_function
+        self.activation_function = activation_function
+        self.hidden_expansion = hidden_expansion
+        # decoder
+        self.d_model = d_model
+        self.num_queries = num_queries
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_in_channels = decoder_in_channels
+        self.num_feature_levels = num_feature_levels
+        self.decoder_n_points = decoder_n_points
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.decoder_activation_function = decoder_activation_function
+        self.attention_dropout = attention_dropout
+        self.num_denoising = num_denoising
+        self.label_noise_ratio = label_noise_ratio
+        self.box_noise_scale = box_noise_scale
+        self.learn_initial_query = learn_initial_query
+        self.anchor_image_size = anchor_image_size
+        self.auxiliary_loss = auxiliary_loss
+        self.disable_custom_kernels = disable_custom_kernels
+        self.with_box_refine = with_box_refine
+        # Loss
+        self.matcher_alpha = matcher_alpha
+        self.matcher_gamma = matcher_gamma
+        self.matcher_class_cost = matcher_class_cost
+        self.matcher_bbox_cost = matcher_bbox_cost
+        self.matcher_giou_cost = matcher_giou_cost
+        self.use_focal_loss = use_focal_loss
+        self.focal_loss_alpha = focal_loss_alpha
+        self.focal_loss_gamma = focal_loss_gamma
+        self.weight_loss_vfl = weight_loss_vfl
+        self.weight_loss_bbox = weight_loss_bbox
+        self.weight_loss_giou = weight_loss_giou
+        self.eos_coefficient = eos_coefficient
+        super().__init__(is_encoder_decoder=is_encoder_decoder, **kwargs)
+
+    @property
+    def num_attention_heads(self) -> int:
+        return self.encoder_attention_heads
+
+    @property
+    def hidden_size(self) -> int:
+        return self.d_model
+
+    @classmethod
+    def from_backbone_configs(cls, backbone_config: PretrainedConfig, **kwargs):
+        """Instantiate a [`RTDetrConfig`] (or a derived class) from a pre-trained backbone model configuration and DETR model
+        configuration.
+
+            Args:
+                backbone_config ([`PretrainedConfig`]):
+                    The backbone configuration.
+
+            Returns:
+                [`RTDetrConfig`]: An instance of a configuration object
+        """
+        return cls(
+            backbone_config=backbone_config,
+            **kwargs,
+        )

src/transformers/models/rt_detr/configuration_rt_detr_resnet.py

+# coding=utf-8
+# Copyright 2024 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.
+"""RT-DETR ResNet model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ...utils.backbone_utils import BackboneConfigMixin, get_aligned_output_features_output_indices
+
+
+logger = logging.get_logger(__name__)
+
+
+class RTDetrResNetConfig(BackboneConfigMixin, PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`RTDetrResnetBackbone`]. It is used to instantiate an
+    ResNet model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the ResNet
+    [microsoft/resnet-50](https://huggingface.co/microsoft/resnet-50) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        embedding_size (`int`, *optional*, defaults to 64):
+            Dimensionality (hidden size) for the embedding layer.
+        hidden_sizes (`List[int]`, *optional*, defaults to `[256, 512, 1024, 2048]`):
+            Dimensionality (hidden size) at each stage.
+        depths (`List[int]`, *optional*, defaults to `[3, 4, 6, 3]`):
+            Depth (number of layers) for each stage.
+        layer_type (`str`, *optional*, defaults to `"bottleneck"`):
+            The layer to use, it can be either `"basic"` (used for smaller models, like resnet-18 or resnet-34) or
+            `"bottleneck"` (used for larger models like resnet-50 and above).
+        hidden_act (`str`, *optional*, defaults to `"relu"`):
+            The non-linear activation function in each block. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"`
+            are supported.
+        downsample_in_first_stage (`bool`, *optional*, defaults to `False`):
+            If `True`, the first stage will downsample the inputs using a `stride` of 2.
+        downsample_in_bottleneck (`bool`, *optional*, defaults to `False`):
+            If `True`, the first conv 1x1 in ResNetBottleNeckLayer will downsample the inputs using a `stride` of 2.
+        out_features (`List[str]`, *optional*):
+            If used as backbone, list of features to output. Can be any of `"stem"`, `"stage1"`, `"stage2"`, etc.
+            (depending on how many stages the model has). If unset and `out_indices` is set, will default to the
+            corresponding stages. If unset and `out_indices` is unset, will default to the last stage. Must be in the
+            same order as defined in the `stage_names` attribute.
+        out_indices (`List[int]`, *optional*):
+            If used as backbone, list of indices of features to output. Can be any of 0, 1, 2, etc. (depending on how
+            many stages the model has). If unset and `out_features` is set, will default to the corresponding stages.
+            If unset and `out_features` is unset, will default to the last stage. Must be in the
+            same order as defined in the `stage_names` attribute.
+
+    Example:
+    ```python
+    >>> from transformers import RTDetrResNetConfig, RTDetrResnetBackbone
+
+    >>> # Initializing a ResNet resnet-50 style configuration
+    >>> configuration = RTDetrResNetConfig()
+
+    >>> # Initializing a model (with random weights) from the resnet-50 style configuration
+    >>> model = RTDetrResnetBackbone(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "rt_detr_resnet"
+    layer_types = ["basic", "bottleneck"]
+
+    def __init__(
+        self,
+        num_channels=3,
+        embedding_size=64,
+        hidden_sizes=[256, 512, 1024, 2048],
+        depths=[3, 4, 6, 3],
+        layer_type="bottleneck",
+        hidden_act="relu",
+        downsample_in_first_stage=False,
+        downsample_in_bottleneck=False,
+        out_features=None,
+        out_indices=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        if layer_type not in self.layer_types:
+            raise ValueError(f"layer_type={layer_type} is not one of {','.join(self.layer_types)}")
+        self.num_channels = num_channels
+        self.embedding_size = embedding_size
+        self.hidden_sizes = hidden_sizes
+        self.depths = depths
+        self.layer_type = layer_type
+        self.hidden_act = hidden_act
+        self.downsample_in_first_stage = downsample_in_first_stage
+        self.downsample_in_bottleneck = downsample_in_bottleneck
+        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, len(depths) + 1)]
+        self._out_features, self._out_indices = get_aligned_output_features_output_indices(
+            out_features=out_features, out_indices=out_indices, stage_names=self.stage_names
+        )

src/transformers/models/rt_detr/modeling_rt_detr_resnet.py

+# coding=utf-8
+# Copyright 2024 Microsoft Research, Inc. and 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.
+"""
+PyTorch RTDetr specific ResNet model. The main difference between hugginface ResNet model is that this RTDetrResNet model forces to use shortcut at the first layer in the resnet-18/34 models.
+See https://github.com/lyuwenyu/RT-DETR/blob/5b628eaa0a2fc25bdafec7e6148d5296b144af85/rtdetr_pytorch/src/nn/backbone/presnet.py#L126 for details.
+"""
+
+from typing import Optional
+
+from torch import Tensor, nn
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BackboneOutput,
+    BaseModelOutputWithNoAttention,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from ...utils.backbone_utils import BackboneMixin
+from .configuration_rt_detr_resnet import RTDetrResNetConfig
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "RTDetrResNetConfig"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "microsoft/resnet-50"
+_EXPECTED_OUTPUT_SHAPE = [1, 2048, 7, 7]
+
+
+# Copied from transformers.models.resnet.modeling_resnet.ResNetConvLayer -> RTDetrResNetConvLayer
+class RTDetrResNetConvLayer(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, kernel_size: int = 3, stride: int = 1, activation: str = "relu"
+    ):
+        super().__init__()
+        self.convolution = nn.Conv2d(
+            in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=kernel_size // 2, bias=False
+        )
+        self.normalization = nn.BatchNorm2d(out_channels)
+        self.activation = ACT2FN[activation] if activation is not None else nn.Identity()
+
+    def forward(self, input: Tensor) -> Tensor:
+        hidden_state = self.convolution(input)
+        hidden_state = self.normalization(hidden_state)
+        hidden_state = self.activation(hidden_state)
+        return hidden_state
+
+
+class RTDetrResNetEmbeddings(nn.Module):
+    """
+    ResNet Embeddings (stem) composed of a deep aggressive convolution.
+    """
+
+    def __init__(self, config: RTDetrResNetConfig):
+        super().__init__()
+        self.embedder = nn.Sequential(
+            *[
+                RTDetrResNetConvLayer(
+                    config.num_channels,
+                    config.embedding_size // 2,
+                    kernel_size=3,
+                    stride=2,
+                    activation=config.hidden_act,
+                ),
+                RTDetrResNetConvLayer(
+                    config.embedding_size // 2,
+                    config.embedding_size // 2,
+                    kernel_size=3,
+                    stride=1,
+                    activation=config.hidden_act,
+                ),
+                RTDetrResNetConvLayer(
+                    config.embedding_size // 2,
+                    config.embedding_size,
+                    kernel_size=3,
+                    stride=1,
+                    activation=config.hidden_act,
+                ),
+            ]
+        )
+        self.pooler = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.num_channels = config.num_channels
+
+    def forward(self, pixel_values: Tensor) -> Tensor:
+        num_channels = pixel_values.shape[1]
+        if num_channels != self.num_channels:
+            raise ValueError(
+                "Make sure that the channel dimension of the pixel values match with the one set in the configuration."
+            )
+        embedding = self.embedder(pixel_values)
+        embedding = self.pooler(embedding)
+        return embedding
+
+
+# Copied from transformers.models.resnet.modeling_resnet.ResNetShortCut -> RTDetrResNetChortCut
+class RTDetrResNetShortCut(nn.Module):
+    """
+    ResNet shortcut, used to project the residual features to the correct size. If needed, it is also used to
+    downsample the input using `stride=2`.
+    """
+
+    def __init__(self, in_channels: int, out_channels: int, stride: int = 2):
+        super().__init__()
+        self.convolution = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False)
+        self.normalization = nn.BatchNorm2d(out_channels)
+
+    def forward(self, input: Tensor) -> Tensor:
+        hidden_state = self.convolution(input)
+        hidden_state = self.normalization(hidden_state)
+        return hidden_state
+
+
+class RTDetrResNetBasicLayer(nn.Module):
+    """
+    A classic ResNet's residual layer composed by two `3x3` convolutions.
+    See https://github.com/lyuwenyu/RT-DETR/blob/5b628eaa0a2fc25bdafec7e6148d5296b144af85/rtdetr_pytorch/src/nn/backbone/presnet.py#L34.
+    """
+
+    def __init__(
+        self,
+        config: RTDetrResNetConfig,
+        in_channels: int,
+        out_channels: int,
+        stride: int = 1,
+        should_apply_shortcut: bool = False,
+    ):
+        super().__init__()
+        if in_channels != out_channels:
+            self.shortcut = (
+                nn.Sequential(
+                    *[nn.AvgPool2d(2, 2, 0, ceil_mode=True), RTDetrResNetShortCut(in_channels, out_channels, stride=1)]
+                )
+                if should_apply_shortcut
+                else nn.Identity()
+            )
+        else:
+            self.shortcut = (
+                RTDetrResNetShortCut(in_channels, out_channels, stride=stride)
+                if should_apply_shortcut
+                else nn.Identity()
+            )
+        self.layer = nn.Sequential(
+            RTDetrResNetConvLayer(in_channels, out_channels, stride=stride),
+            RTDetrResNetConvLayer(out_channels, out_channels, activation=None),
+        )
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_state):
+        residual = hidden_state
+        hidden_state = self.layer(hidden_state)
+        residual = self.shortcut(residual)
+        hidden_state += residual
+        hidden_state = self.activation(hidden_state)
+        return hidden_state
+
+
+class RTDetrResNetBottleNeckLayer(nn.Module):
+    """
+    A classic RTDetrResNet's bottleneck layer composed by three `3x3` convolutions.
+
+    The first `1x1` convolution reduces the input by a factor of `reduction` in order to make the second `3x3`
+    convolution faster. The last `1x1` convolution remaps the reduced features to `out_channels`. If
+    `downsample_in_bottleneck` is true, downsample will be in the first layer instead of the second layer.
+    """
+
+    def __init__(
+        self,
+        config: RTDetrResNetConfig,
+        in_channels: int,
+        out_channels: int,
+        stride: int = 1,
+    ):
+        super().__init__()
+        reduction = 4
+        should_apply_shortcut = in_channels != out_channels or stride != 1
+        reduces_channels = out_channels // reduction
+        if stride == 2:
+            self.shortcut = nn.Sequential(
+                *[
+                    nn.AvgPool2d(2, 2, 0, ceil_mode=True),
+                    RTDetrResNetShortCut(in_channels, out_channels, stride=1)
+                    if should_apply_shortcut
+                    else nn.Identity(),
+                ]
+            )
+        else:
+            self.shortcut = (
+                RTDetrResNetShortCut(in_channels, out_channels, stride=stride)
+                if should_apply_shortcut
+                else nn.Identity()
+            )
+        self.layer = nn.Sequential(
+            RTDetrResNetConvLayer(
+                in_channels, reduces_channels, kernel_size=1, stride=stride if config.downsample_in_bottleneck else 1
+            ),
+            RTDetrResNetConvLayer(
+                reduces_channels, reduces_channels, stride=stride if not config.downsample_in_bottleneck else 1
+            ),
+            RTDetrResNetConvLayer(reduces_channels, out_channels, kernel_size=1, activation=None),
+        )
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_state):
+        residual = hidden_state
+        hidden_state = self.layer(hidden_state)
+        residual = self.shortcut(residual)
+        hidden_state += residual
+        hidden_state = self.activation(hidden_state)
+        return hidden_state
+
+
+class RTDetrResNetStage(nn.Module):
+    """
+    A RTDetrResNet stage composed by stacked layers.
+    """
+
+    def __init__(
+        self,
+        config: RTDetrResNetConfig,
+        in_channels: int,
+        out_channels: int,
+        stride: int = 2,
+        depth: int = 2,
+    ):
+        super().__init__()
+
+        layer = RTDetrResNetBottleNeckLayer if config.layer_type == "bottleneck" else RTDetrResNetBasicLayer
+
+        if config.layer_type == "bottleneck":
+            first_layer = layer(
+                config,
+                in_channels,
+                out_channels,
+                stride=stride,
+            )
+        else:
+            first_layer = layer(config, in_channels, out_channels, stride=stride, should_apply_shortcut=True)
+        self.layers = nn.Sequential(
+            first_layer, *[layer(config, out_channels, out_channels) for _ in range(depth - 1)]
+        )
+
+    def forward(self, input: Tensor) -> Tensor:
+        hidden_state = input
+        for layer in self.layers:
+            hidden_state = layer(hidden_state)
+        return hidden_state
+
+
+# Copied from transformers.models.resnet.modeling_resnet.ResNetEncoder with ResNet->RTDetrResNet
+class RTDetrResNetEncoder(nn.Module):
+    def __init__(self, config: RTDetrResNetConfig):
+        super().__init__()
+        self.stages = nn.ModuleList([])
+        # based on `downsample_in_first_stage` the first layer of the first stage may or may not downsample the input
+        self.stages.append(
+            RTDetrResNetStage(
+                config,
+                config.embedding_size,
+                config.hidden_sizes[0],
+                stride=2 if config.downsample_in_first_stage else 1,
+                depth=config.depths[0],
+            )
+        )
+        in_out_channels = zip(config.hidden_sizes, config.hidden_sizes[1:])
+        for (in_channels, out_channels), depth in zip(in_out_channels, config.depths[1:]):
+            self.stages.append(RTDetrResNetStage(config, in_channels, out_channels, depth=depth))
+
+    def forward(
+        self, hidden_state: Tensor, output_hidden_states: bool = False, return_dict: bool = True
+    ) -> BaseModelOutputWithNoAttention:
+        hidden_states = () if output_hidden_states else None
+
+        for stage_module in self.stages:
+            if output_hidden_states:
+                hidden_states = hidden_states + (hidden_state,)
+
+            hidden_state = stage_module(hidden_state)
+
+        if output_hidden_states:
+            hidden_states = hidden_states + (hidden_state,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_state, hidden_states] if v is not None)
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=hidden_state,
+            hidden_states=hidden_states,
+        )
+
+
+# Copied from transformers.models.resnet.modeling_resnet.ResNetPreTrainedModel with ResNet->RTDetrResNet
+class RTDetrResNetPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = RTDetrResNetConfig
+    base_model_prefix = "resnet"
+    main_input_name = "pixel_values"
+    _no_split_modules = ["RTDetrResNetConvLayer", "RTDetrResNetShortCut"]
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Conv2d):
+            nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
+        elif isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
+            nn.init.constant_(module.weight, 1)
+            nn.init.constant_(module.bias, 0)
+
+
+RTDETR_RESNET_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it
+    as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`RTDetrResNetConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+RTDETR_RESNET_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoImageProcessor`]. See
+            [`RTDetrImageProcessor.__call__`] for details.
+
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    """
+    ResNet backbone, to be used with frameworks like RTDETR.
+    """,
+    RTDETR_RESNET_START_DOCSTRING,
+)
+class RTDetrResNetBackbone(RTDetrResNetPreTrainedModel, BackboneMixin):
+    def __init__(self, config):
+        super().__init__(config)
+        super()._init_backbone(config)
+
+        self.num_features = [config.embedding_size] + config.hidden_sizes
+        self.embedder = RTDetrResNetEmbeddings(config)
+        self.encoder = RTDetrResNetEncoder(config)
+
+        # initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(RTDETR_RESNET_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BackboneOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self, pixel_values: Tensor, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None
+    ) -> BackboneOutput:
+        """
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import RTDetrResNetConfig, RTDetrResNetBackbone
+        >>> import torch
+
+        >>> config = RTDetrResNetConfig()
+        >>> model = RTDetrResNetBackbone(config)
+
+        >>> pixel_values = torch.randn(1, 3, 224, 224)
+
+        >>> with torch.no_grad():
+        ...     outputs = model(pixel_values)
+
+        >>> feature_maps = outputs.feature_maps
+        >>> list(feature_maps[-1].shape)
+        [1, 2048, 7, 7]
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        embedding_output = self.embedder(pixel_values)
+
+        outputs = self.encoder(embedding_output, output_hidden_states=True, return_dict=True)
+
+        hidden_states = outputs.hidden_states
+
+        feature_maps = ()
+        for idx, stage in enumerate(self.stage_names):
+            if stage in self.out_features:
+                feature_maps += (hidden_states[idx],)
+
+        if not return_dict:
+            output = (feature_maps,)
+            if output_hidden_states:
+                output += (outputs.hidden_states,)
+            return output
+
+        return BackboneOutput(
+            feature_maps=feature_maps,
+            hidden_states=outputs.hidden_states if output_hidden_states else None,
+            attentions=None,
+        )

src/transformers/models/timm_backbone/modeling_timm_backbone.py

@@ -113,10 +113,10 @@ class TimmBackbone(PreTrainedModel, BackboneMixin):
         return super()._from_config(config, **kwargs)
 
     def freeze_batch_norm_2d(self):
-        timm.layers.freeze_batch_norm_2d(self._backbone)
+        timm.utils.model.freeze_batch_norm_2d(self._backbone)
 
     def unfreeze_batch_norm_2d(self):
-        timm.layers.unfreeze_batch_norm_2d(self._backbone)
+        timm.utils.model.unfreeze_batch_norm_2d(self._backbone)
 
     def _init_weights(self, module):
         """

src/transformers/utils/backbone_utils.py

@@ -313,7 +313,6 @@ def load_backbone(config):
     use_pretrained_backbone = getattr(config, "use_pretrained_backbone", None)
     backbone_checkpoint = getattr(config, "backbone", None)
     backbone_kwargs = getattr(config, "backbone_kwargs", None)
-
     backbone_kwargs = {} if backbone_kwargs is None else backbone_kwargs
 
     if backbone_kwargs and backbone_config is not None:

src/transformers/utils/dummy_pt_objects.py

@@ -7492,6 +7492,41 @@ def load_tf_weights_in_roformer(*args, **kwargs):
     requires_backends(load_tf_weights_in_roformer, ["torch"])
 
 
+class RTDetrForObjectDetection(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class RTDetrModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class RTDetrPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class RTDetrResNetBackbone(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class RTDetrResNetPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class RwkvForCausalLM(metaclass=DummyObject):
     _backends = ["torch"]
 

src/transformers/utils/dummy_vision_objects.py

@@ -492,6 +492,13 @@ class PvtImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class RTDetrImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class SamImageProcessor(metaclass=DummyObject):
     _backends = ["vision"]