Add Deformable DETR (#17281)

* First draft

* More improvements

* Improve model, add custom CUDA code

* Import torch before

* Add script that imports custom layer

* Add everything in new ops directory

* Import custom layer in modeling file

* Fix ARCHIVE_MAP typo

* Creating the custom kernel on the fly.

* Import custom layer in modeling file

* More improvements

* Fix CUDA loading

* More improvements

* Improve conversion script

* Improve conversion script

* Make it work until encoder_outputs

* Make forward pass work

* More improvements

* Make logits match original implementation

* Make implementation also support single_scale model

* Add support for single_scale and dilation checkpoint

* Add support for with_box_refine model

* Support also two stage model

* Improve tests

* Fix more tests

* Make more tests pass

* Upload all models to the hub

* Clean up some code

* Improve decoder outputs

* Rename intermediate hidden states and reference points

* Improve model outputs

* Move tests to dedicated folder

* Improve model outputs

* Fix retain_grad test

* Improve docs

* Clean up and make test_initialization pass

* Improve variable names

* Add copied from statements

* Improve docs

* Fix style

* Improve docs

* Improve docs, move tests to model folder

* Fix rebase

* Remove DetrForSegmentation from auto mapping

* Apply suggestions from code review

* Improve variable names and docstrings

* Apply some more suggestions from code review

* Apply suggestion from code review

* better docs and variables names

* hint to num_queries and two_stage confusion

* remove asserts and code refactor

* add exception if two_stage is True and with_box_refine is False

* use f-strings

* Improve docs and variable names

* Fix code quality

* Fix rebase

* Add require_torch_gpu decorator

* Add pip install ninja to CI jobs

* Apply suggestion of @sgugger

* Remove DeformableDetrForObjectDetection from auto mapping

* Remove DeformableDetrModel from auto mapping

* Add model to toctree

* Add model back to mappings, skip model in pipeline tests

* Apply @sgugger's suggestion

* Fix imports in the init

* Fix copies

* Add CPU implementation

* Comment out GPU function

* Undo previous change

* Apply more suggestions

* Remove require_torch_gpu annotator

* Fix quality

* Add logger.info

* Fix logger

* Fix variable names

* Fix initializaztion

* Add missing initialization

* Update checkpoint name

* Add model to doc tests

* Add CPU/GPU equivalence test

* Add Deformable DETR to pipeline tests

* Skip model for object detection pipeline
Co-authored-by: Nicolas Patry <patry.nicolas@protonmail.com>
Co-authored-by: Nouamane Tazi <nouamane98@gmail.com>
Co-authored-by: Sylvain Gugger <Sylvain.gugger@gmail.com>

README.md

@@ -285,6 +285,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[Decision Transformer](https://huggingface.co/docs/transformers/model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
+1. **[Deformable DETR](https://huggingface.co/docs/transformers/main/model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.

README_ko.md

@@ -237,6 +237,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[Decision Transformer](https://huggingface.co/docs/transformers/model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
+1. **[Deformable DETR](https://huggingface.co/docs/transformers/main/model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.

README_zh-hans.md

@@ -261,6 +261,7 @@ conda install -c huggingface transformers
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (来自 Microsoft) 伴随论文 [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) 由 Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen 发布。
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (来自 Microsoft) 伴随论文 [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) 由 Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen 发布。
 1. **[Decision Transformer](https://huggingface.co/docs/transformers/model_doc/decision_transformer)** (来自 Berkeley/Facebook/Google) 伴随论文 [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) 由 Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch 发布。
+1. **[Deformable DETR](https://huggingface.co/docs/transformers/main/model_doc/deformable_detr)** (来自 SenseTime Research) 伴随论文 [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) 由 Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai 发布。
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (来自 Facebook) 伴随论文 [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) 由 Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou 发布。
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (来自 Facebook) 伴随论文 [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) 由 Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko 发布。
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (来自 Microsoft Research) 伴随论文 [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) 由 Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan 发布。

README_zh-hant.md

@@ -273,6 +273,7 @@ conda install -c huggingface transformers
 1. **[DeBERTa](https://huggingface.co/docs/transformers/model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](https://huggingface.co/docs/transformers/model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[Decision Transformer](https://huggingface.co/docs/transformers/model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
+1. **[Deformable DETR](https://huggingface.co/docs/transformers/main/model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
 1. **[DeiT](https://huggingface.co/docs/transformers/model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.

docs/source/en/_toctree.yml

@@ -364,6 +364,8 @@
         title: ConvNeXT
       - local: model_doc/cvt
         title: CvT
+      - local: model_doc/deformable_detr
+        title: Deformable DETR
       - local: model_doc/deit
         title: DeiT
       - local: model_doc/detr

docs/source/en/index.mdx

@@ -77,6 +77,7 @@ The documentation is organized into five sections:
 1. **[DeBERTa](model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DeBERTa-v2](model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[Decision Transformer](model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
+1. **[Deformable DETR](model_doc/deformable_detr)** (from SenseTime Research) released with the paper [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
 1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
 1. **[DETR](model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
@@ -223,6 +224,7 @@ Flax), PyTorch, and/or TensorFlow.
 |           DeBERTa           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |         DeBERTa-v2          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |    Decision Transformer     |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|       Deformable DETR       |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            DeiT             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            DETR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         DistilBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/deformable_detr.mdx

+<!--Copyright 2022 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.
+-->
+
+# Deformable DETR
+
+## Overview
+
+The Deformable DETR model was proposed in [Deformable DETR: Deformable Transformers for End-to-End Object Detection](https://arxiv.org/abs/2010.04159) by Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, Jifeng Dai.
+Deformable DETR mitigates the slow convergence issues and limited feature spatial resolution of the original [DETR](detr) by leveraging a new deformable attention module which only attends to a small set of key sampling points around a reference.
+
+The abstract from the paper is the following:
+
+*DETR has been recently proposed to eliminate the need for many hand-designed components in object detection while demonstrating good performance. However, it suffers from slow convergence and limited feature spatial resolution, due to the limitation of Transformer attention modules in processing image feature maps. To mitigate these issues, we proposed Deformable DETR, whose attention modules only attend to a small set of key sampling points around a reference. Deformable DETR can achieve better performance than DETR (especially on small objects) with 10 times less training epochs. Extensive experiments on the COCO benchmark demonstrate the effectiveness of our approach.*
+
+Tips:
+
+- One can use the [`AutoFeatureExtractor`] API to prepare images (and optional targets) for the model. This will instantiate a [`DetrFeatureExtractor`] behind the scenes.
+- Training Deformable DETR is equivalent to training the original [DETR](detr) model. Demo notebooks can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/DETR).
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/deformable_detr_architecture.png"
+alt="drawing" width="600"/> 
+
+<small> Deformable DETR architecture. Taken from the <a href="https://arxiv.org/abs/2010.04159">original paper</a>.</small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code can be found [here](https://github.com/fundamentalvision/Deformable-DETR).
+
+## DeformableDetrConfig
+
+[[autodoc]] DeformableDetrConfig
+
+
+## DeformableDetrModel
+
+[[autodoc]] DeformableDetrModel
+    - forward
+
+
+## DeformableDetrForObjectDetection
+
+[[autodoc]] DeformableDetrForObjectDetection
+    - forward
\ No newline at end of file

setup.py

@@ -411,6 +411,7 @@ setup(
     url="https://github.com/huggingface/transformers",
     package_dir={"": "src"},
     packages=find_packages("src"),
+    package_data={"transformers": ["py.typed", "*.cu", "*.cpp", "*.cuh", "*.h"]},
     zip_safe=False,
     extras_require=extras,
     entry_points={"console_scripts": ["transformers-cli=transformers.commands.transformers_cli:main"]},

src/transformers/__init__.py

@@ -187,6 +187,7 @@ _import_structure = {
     "models.deberta": ["DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaConfig", "DebertaTokenizer"],
     "models.deberta_v2": ["DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaV2Config"],
     "models.decision_transformer": ["DECISION_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "DecisionTransformerConfig"],
+    "models.deformable_detr": ["DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DeformableDetrConfig"],
     "models.deit": ["DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DeiTConfig"],
     "models.detr": ["DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DetrConfig"],
     "models.dialogpt": [],
@@ -682,12 +683,20 @@ try:
     if not (is_timm_available() and is_vision_available()):
         raise OptionalDependencyNotAvailable()
 except OptionalDependencyNotAvailable:
-    from .utils import dummy_timm_objects
+    from .utils import dummy_timm_and_vision_objects
 
-    _import_structure["utils.dummy_timm_objects"] = [
-        name for name in dir(dummy_timm_objects) if not name.startswith("_")
+    _import_structure["utils.dummy_timm_and_vision_objects"] = [
+        name for name in dir(dummy_timm_and_vision_objects) if not name.startswith("_")
     ]
 else:
+    _import_structure["models.deformable_detr"].extend(
+        [
+            "DEFORMABLE_DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "DeformableDetrForObjectDetection",
+            "DeformableDetrModel",
+            "DeformableDetrPreTrainedModel",
+        ]
+    )
     _import_structure["models.detr"].extend(
         [
             "DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3072,6 +3081,7 @@ if TYPE_CHECKING:
         DECISION_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
         DecisionTransformerConfig,
     )
+    from .models.deformable_detr import DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DeformableDetrConfig
     from .models.deit import DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, DeiTConfig
     from .models.detr import DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DetrConfig
     from .models.distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig, DistilBertTokenizer
@@ -3502,8 +3512,14 @@ if TYPE_CHECKING:
         if not (is_timm_available() and is_vision_available()):
             raise OptionalDependencyNotAvailable()
     except OptionalDependencyNotAvailable:
-        from .utils.dummy_timm_objects import *
+        from .utils.dummy_timm_and_vision_objects import *
     else:
+        from .models.deformable_detr import (
+            DEFORMABLE_DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DeformableDetrForObjectDetection,
+            DeformableDetrModel,
+            DeformableDetrPreTrainedModel,
+        )
         from .models.detr import (
             DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
             DetrForObjectDetection,

src/transformers/models/__init__.py

@@ -47,6 +47,7 @@ from . import (
     deberta,
     deberta_v2,
     decision_transformer,
+    deformable_detr,
     deit,
     detr,
     dialogpt,

src/transformers/models/auto/configuration_auto.py

@@ -53,6 +53,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("deberta", "DebertaConfig"),
         ("deberta-v2", "DebertaV2Config"),
         ("decision_transformer", "DecisionTransformerConfig"),
+        ("deformable_detr", "DeformableDetrConfig"),
         ("deit", "DeiTConfig"),
         ("detr", "DetrConfig"),
         ("distilbert", "DistilBertConfig"),
@@ -182,6 +183,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("data2vec-vision", "DATA2VEC_VISION_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("deberta", "DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("deberta-v2", "DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("deformable_detr", "DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("deit", "DEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("detr", "DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("distilbert", "DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -307,6 +309,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("deberta", "DeBERTa"),
         ("deberta-v2", "DeBERTa-v2"),
         ("decision_transformer", "Decision Transformer"),
+        ("deformable_detr", "Deformable DETR"),
         ("deit", "DeiT"),
         ("detr", "DETR"),
         ("dialogpt", "DialoGPT"),

src/transformers/models/auto/feature_extraction_auto.py

@@ -43,6 +43,7 @@ FEATURE_EXTRACTOR_MAPPING_NAMES = OrderedDict(
         ("cvt", "ConvNextFeatureExtractor"),
         ("data2vec-audio", "Wav2Vec2FeatureExtractor"),
         ("data2vec-vision", "BeitFeatureExtractor"),
+        ("deformable_detr", "DetrFeatureExtractor"),
         ("deit", "DeiTFeatureExtractor"),
         ("detr", "DetrFeatureExtractor"),
         ("detr", "DetrFeatureExtractor"),

src/transformers/models/auto/modeling_auto.py

@@ -53,6 +53,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("deberta-v2", "DebertaV2Model"),
         ("decision_transformer", "DecisionTransformerModel"),
         ("decision_transformer_gpt2", "DecisionTransformerGPT2Model"),
+        ("deformable_detr", "DeformableDetrModel"),
         ("deit", "DeiTModel"),
         ("detr", "DetrModel"),
         ("distilbert", "DistilBertModel"),
@@ -451,6 +452,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Object Detection mapping
+        ("deformable_detr", "DeformableDetrForObjectDetection"),
         ("detr", "DetrForObjectDetection"),
         ("yolos", "YolosForObjectDetection"),
     ]

src/transformers/models/deformable_detr/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 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_timm_available
+
+
+_import_structure = {
+    "configuration_deformable_detr": ["DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DeformableDetrConfig"],
+}
+
+try:
+    if not is_timm_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_deformable_detr"] = [
+        "DEFORMABLE_DETR_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DeformableDetrForObjectDetection",
+        "DeformableDetrModel",
+        "DeformableDetrPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_deformable_detr import DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DeformableDetrConfig
+
+    try:
+        if not is_timm_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_deformable_detr import (
+            DEFORMABLE_DETR_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DeformableDetrForObjectDetection,
+            DeformableDetrModel,
+            DeformableDetrPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/deformable_detr/configuration_deformable_detr.py

+# coding=utf-8
+# Copyright 2022 SenseTime 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.
+""" Deformable DETR model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DEFORMABLE_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "SenseTime/deformable-detr": "https://huggingface.co/sensetime/deformable-detr/resolve/main/config.json",
+    # See all Deformable DETR models at https://huggingface.co/models?filter=deformable-detr
+}
+
+
+class DeformableDetrConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeformableDetrModel`]. It is used to instantiate
+    a Deformable 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 Deformable DETR
+    [SenseTime/deformable-detr](https://huggingface.co/SenseTime/deformable-detr) 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_queries (`int`, *optional*, defaults to 300):
+            Number of object queries, i.e. detection slots. This is the maximal number of objects
+            [`DeformableDetrModel`] can detect in a single image. In case `two_stage` is set to `True`, we use
+            `two_stage_num_proposals` instead.
+        d_model (`int`, *optional*, defaults to 256):
+            Dimension of the layers.
+        encoder_layers (`int`, *optional*, defaults to 6):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 6):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 8):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        init_xavier_std (`float`, *optional*, defaults to 1):
+            The scaling factor used for the Xavier initialization gain in the HM Attention map module.
+        encoder_layerdrop: (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop: (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        auxiliary_loss (`bool`, *optional*, defaults to `False`):
+            Whether auxiliary decoding losses (loss at each decoder layer) are to be used.
+        position_embedding_type (`str`, *optional*, defaults to `"sine"`):
+            Type of position embeddings to be used on top of the image features. One of `"sine"` or `"learned"`.
+        backbone (`str`, *optional*, defaults to `"resnet50"`):
+            Name of convolutional backbone to use. Supports any convolutional backbone from the timm package. For a
+            list of all available models, see [this
+            page](https://rwightman.github.io/pytorch-image-models/#load-a-pretrained-model).
+        dilation (`bool`, *optional*, defaults to `False`):
+            Whether to replace stride with dilation in the last convolutional block (DC5).
+        class_cost (`float`, *optional*, defaults to 1):
+            Relative weight of the classification error in the Hungarian matching cost.
+        bbox_cost (`float`, *optional*, defaults to 5):
+            Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.
+        giou_cost (`float`, *optional*, defaults to 2):
+            Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.
+        mask_loss_coefficient (`float`, *optional*, defaults to 1):
+            Relative weight of the Focal loss in the panoptic segmentation loss.
+        dice_loss_coefficient (`float`, *optional*, defaults to 1):
+            Relative weight of the DICE/F-1 loss in the panoptic segmentation loss.
+        bbox_loss_coefficient (`float`, *optional*, defaults to 5):
+            Relative weight of the L1 bounding box loss in the object detection loss.
+        giou_loss_coefficient (`float`, *optional*, defaults to 2):
+            Relative weight of the generalized IoU loss in the object detection loss.
+        eos_coefficient (`float`, *optional*, defaults to 0.1):
+            Relative classification weight of the 'no-object' class in the object detection loss.
+        num_feature_levels (`int`, *optional*, defaults to 4):
+            The number of input feature levels.
+        encoder_n_points (`int`, *optional*, defaults to 4):
+            The number of sampled keys in each feature level for each attention head in the encoder.
+        decoder_n_points (`int`, *optional*, defaults to 4):
+            The number of sampled keys in each feature level for each attention head in the decoder.
+        two_stage (`bool`, *optional*, defaults to `False`):
+            Whether to apply a two-stage deformable DETR, where the region proposals are also generated by a variant of
+            Deformable DETR, which are further fed into the decoder for iterative bounding box refinement.
+        two_stage_num_proposals (`int`, *optional*, defaults to 300):
+            The number of region proposals to be generated, in case `two_stage` is set to `True`.
+        with_box_refine (`bool`, *optional*, defaults to `False`):
+            Whether to apply iterative bounding box refinement, where each decoder layer refines the bounding boxes
+            based on the predictions from the previous layer.
+
+    Examples:
+
+    ```python
+    >>> from transformers import DeformableDetrModel, DeformableDetrConfig
+
+    >>> # Initializing a Deformable DETR SenseTime/deformable-detr style configuration
+    >>> configuration = DeformableDetrConfig()
+
+    >>> # Initializing a model from the SenseTime/deformable-detr style configuration
+    >>> model = DeformableDetrModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "deformable_detr"
+    attribute_map = {
+        "hidden_size": "d_model",
+        "num_attention_heads": "encoder_attention_heads",
+    }
+
+    def __init__(
+        self,
+        num_queries=300,
+        max_position_embeddings=1024,
+        encoder_layers=6,
+        encoder_ffn_dim=1024,
+        encoder_attention_heads=8,
+        decoder_layers=6,
+        decoder_ffn_dim=1024,
+        decoder_attention_heads=8,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        is_encoder_decoder=True,
+        activation_function="relu",
+        d_model=256,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        init_xavier_std=1.0,
+        return_intermediate=True,
+        auxiliary_loss=False,
+        position_embedding_type="sine",
+        backbone="resnet50",
+        dilation=False,
+        num_feature_levels=4,
+        encoder_n_points=4,
+        decoder_n_points=4,
+        two_stage=False,
+        two_stage_num_proposals=300,
+        with_box_refine=False,
+        class_cost=1,
+        bbox_cost=5,
+        giou_cost=2,
+        mask_loss_coefficient=1,
+        dice_loss_coefficient=1,
+        bbox_loss_coefficient=5,
+        giou_loss_coefficient=2,
+        eos_coefficient=0.1,
+        **kwargs
+    ):
+        self.num_queries = num_queries
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.init_xavier_std = init_xavier_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.auxiliary_loss = auxiliary_loss
+        self.position_embedding_type = position_embedding_type
+        self.backbone = backbone
+        self.dilation = dilation
+        # deformable attributes
+        self.num_feature_levels = num_feature_levels
+        self.encoder_n_points = encoder_n_points
+        self.decoder_n_points = decoder_n_points
+        self.two_stage = two_stage
+        self.two_stage_num_proposals = two_stage_num_proposals
+        self.with_box_refine = with_box_refine
+        if two_stage is True and with_box_refine is False:
+            raise ValueError("If two_stage is True, with_box_refine must be True.")
+        # Hungarian matcher
+        self.class_cost = class_cost
+        self.bbox_cost = bbox_cost
+        self.giou_cost = giou_cost
+        # Loss coefficients
+        self.mask_loss_coefficient = mask_loss_coefficient
+        self.dice_loss_coefficient = dice_loss_coefficient
+        self.bbox_loss_coefficient = bbox_loss_coefficient
+        self.giou_loss_coefficient = giou_loss_coefficient
+        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

src/transformers/models/deformable_detr/convert_deformable_detr_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# 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.
+"""Convert Deformable DETR checkpoints."""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+
+import requests
+from huggingface_hub import cached_download, hf_hub_url
+from transformers import DeformableDetrConfig, DeformableDetrForObjectDetection, DetrFeatureExtractor
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def rename_key(orig_key):
+    if "backbone.0.body" in orig_key:
+        orig_key = orig_key.replace("backbone.0.body", "backbone.conv_encoder.model")
+    if "transformer" in orig_key:
+        orig_key = orig_key.replace("transformer.", "")
+    if "norm1" in orig_key:
+        if "encoder" in orig_key:
+            orig_key = orig_key.replace("norm1", "self_attn_layer_norm")
+        else:
+            orig_key = orig_key.replace("norm1", "encoder_attn_layer_norm")
+    if "norm2" in orig_key:
+        if "encoder" in orig_key:
+            orig_key = orig_key.replace("norm2", "final_layer_norm")
+        else:
+            orig_key = orig_key.replace("norm2", "self_attn_layer_norm")
+    if "norm3" in orig_key:
+        orig_key = orig_key.replace("norm3", "final_layer_norm")
+    if "linear1" in orig_key:
+        orig_key = orig_key.replace("linear1", "fc1")
+    if "linear2" in orig_key:
+        orig_key = orig_key.replace("linear2", "fc2")
+    if "query_embed" in orig_key:
+        orig_key = orig_key.replace("query_embed", "query_position_embeddings")
+    if "cross_attn" in orig_key:
+        orig_key = orig_key.replace("cross_attn", "encoder_attn")
+
+    return orig_key
+
+
+def read_in_q_k_v(state_dict):
+    # transformer decoder self-attention layers
+    for i in range(6):
+        # read in weights + bias of input projection layer of self-attention
+        in_proj_weight = state_dict.pop(f"decoder.layers.{i}.self_attn.in_proj_weight")
+        in_proj_bias = state_dict.pop(f"decoder.layers.{i}.self_attn.in_proj_bias")
+        # next, add query, keys and values (in that order) to the state dict
+        state_dict[f"decoder.layers.{i}.self_attn.q_proj.weight"] = in_proj_weight[:256, :]
+        state_dict[f"decoder.layers.{i}.self_attn.q_proj.bias"] = in_proj_bias[:256]
+        state_dict[f"decoder.layers.{i}.self_attn.k_proj.weight"] = in_proj_weight[256:512, :]
+        state_dict[f"decoder.layers.{i}.self_attn.k_proj.bias"] = in_proj_bias[256:512]
+        state_dict[f"decoder.layers.{i}.self_attn.v_proj.weight"] = in_proj_weight[-256:, :]
+        state_dict[f"decoder.layers.{i}.self_attn.v_proj.bias"] = in_proj_bias[-256:]
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+
+    return im
+
+
+@torch.no_grad()
+def convert_deformable_detr_checkpoint(
+    checkpoint_path,
+    single_scale,
+    dilation,
+    with_box_refine,
+    two_stage,
+    pytorch_dump_folder_path,
+    push_to_hub,
+):
+    """
+    Copy/paste/tweak model's weights to our Deformable DETR structure.
+    """
+
+    # load default config
+    config = DeformableDetrConfig()
+    # set config attributes
+    if single_scale:
+        config.num_feature_levels = 1
+    config.dilation = dilation
+    config.with_box_refine = with_box_refine
+    config.two_stage = two_stage
+    # set labels
+    config.num_labels = 91
+    repo_id = "datasets/huggingface/label-files"
+    filename = "coco-detection-id2label.json"
+    id2label = json.load(open(cached_download(hf_hub_url(repo_id, filename)), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    config.id2label = id2label
+    config.label2id = {v: k for k, v in id2label.items()}
+
+    # load feature extractor
+    feature_extractor = DetrFeatureExtractor(format="coco_detection")
+
+    # prepare image
+    img = prepare_img()
+    encoding = feature_extractor(images=img, return_tensors="pt")
+    pixel_values = encoding["pixel_values"]
+
+    logger.info("Converting model...")
+
+    # load original state dict
+    state_dict = torch.load(checkpoint_path, map_location="cpu")["model"]
+    # rename keys
+    for key in state_dict.copy().keys():
+        val = state_dict.pop(key)
+        state_dict[rename_key(key)] = val
+    # query, key and value matrices need special treatment
+    read_in_q_k_v(state_dict)
+    # important: we need to prepend a prefix to each of the base model keys as the head models use different attributes for them
+    prefix = "model."
+    for key in state_dict.copy().keys():
+        if not key.startswith("class_embed") and not key.startswith("bbox_embed"):
+            val = state_dict.pop(key)
+            state_dict[prefix + key] = val
+    # finally, create HuggingFace model and load state dict
+    model = DeformableDetrForObjectDetection(config)
+    model.load_state_dict(state_dict)
+    model.eval()
+
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    model.to(device)
+    # verify our conversion
+    outputs = model(pixel_values.to(device))
+
+    expected_logits = torch.tensor(
+        [[-9.6645, -4.3449, -5.8705], [-9.7035, -3.8504, -5.0724], [-10.5634, -5.3379, -7.5116]]
+    )
+    expected_boxes = torch.tensor([[0.8693, 0.2289, 0.2492], [0.3150, 0.5489, 0.5845], [0.5563, 0.7580, 0.8518]])
+
+    if single_scale:
+        expected_logits = torch.tensor(
+            [[-9.9051, -4.2541, -6.4852], [-9.6947, -4.0854, -6.8033], [-10.0665, -5.8470, -7.7003]]
+        )
+        expected_boxes = torch.tensor([[0.7292, 0.4991, 0.5532], [0.7959, 0.2426, 0.4236], [0.7582, 0.3518, 0.4451]])
+
+    if single_scale and dilation:
+        expected_logits = torch.tensor(
+            [[-8.9652, -4.1074, -5.6635], [-9.0596, -4.9447, -6.6075], [-10.1178, -4.5275, -6.2671]]
+        )
+        expected_boxes = torch.tensor([[0.7665, 0.4130, 0.4769], [0.8364, 0.1841, 0.3391], [0.6261, 0.3895, 0.7978]])
+
+    if with_box_refine:
+        expected_logits = torch.tensor(
+            [[-8.8895, -5.4187, -6.8153], [-8.4706, -6.1668, -7.6184], [-9.0042, -5.5359, -6.9141]]
+        )
+        expected_boxes = torch.tensor([[0.7828, 0.2208, 0.4323], [0.0892, 0.5996, 0.1319], [0.5524, 0.6389, 0.8914]])
+
+    if with_box_refine and two_stage:
+        expected_logits = torch.tensor(
+            [[-6.7108, -4.3213, -6.3777], [-8.9014, -6.1799, -6.7240], [-6.9315, -4.4735, -6.2298]]
+        )
+        expected_boxes = torch.tensor([[0.2583, 0.5499, 0.4683], [0.7652, 0.9068, 0.4882], [0.5490, 0.2763, 0.0564]])
+
+    print("Logits:", outputs.logits[0, :3, :3])
+
+    assert torch.allclose(outputs.logits[0, :3, :3], expected_logits.to(device), atol=1e-4)
+    assert torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes.to(device), atol=1e-4)
+
+    print("Everything ok!")
+
+    # Save model and feature extractor
+    logger.info(f"Saving PyTorch model and feature extractor to {pytorch_dump_folder_path}...")
+    Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+    model.save_pretrained(pytorch_dump_folder_path)
+    feature_extractor.save_pretrained(pytorch_dump_folder_path)
+
+    # Push to hub
+    if push_to_hub:
+        model_name = "deformable-detr"
+        model_name += "-single-scale" if single_scale else ""
+        model_name += "-dc5" if dilation else ""
+        model_name += "-with-box-refine" if with_box_refine else ""
+        model_name += "-two-stage" if two_stage else ""
+        print("Pushing model to hub...")
+        model.push_to_hub(repo_path_or_name=model_name, organization="nielsr", commit_message="Add model")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--checkpoint_path",
+        type=str,
+        default="/home/niels/checkpoints/deformable_detr/r50_deformable_detr-checkpoint.pth",
+        help="Path to Pytorch checkpoint (.pth file) you'd like to convert.",
+    )
+    parser.add_argument("--single_scale", action="store_true", help="Whether to set config.num_features_levels = 1.")
+    parser.add_argument("--dilation", action="store_true", help="Whether to set config.dilation=True.")
+    parser.add_argument("--with_box_refine", action="store_true", help="Whether to set config.with_box_refine=True.")
+    parser.add_argument("--two_stage", action="store_true", help="Whether to set config.two_stage=True.")
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to the folder to output PyTorch model.",
+    )
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+    args = parser.parse_args()
+    convert_deformable_detr_checkpoint(
+        args.checkpoint_path,
+        args.single_scale,
+        args.dilation,
+        args.with_box_refine,
+        args.two_stage,
+        args.pytorch_dump_folder_path,
+        args.push_to_hub,
+    )

src/transformers/models/deformable_detr/custom_kernel/cpu/ms_deform_attn_cpu.cpp

+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#include <vector>
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+
+
+at::Tensor
+ms_deform_attn_cpu_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step)
+{
+    AT_ERROR("Not implement on cpu");
+}
+
+std::vector<at::Tensor>
+ms_deform_attn_cpu_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step)
+{
+    AT_ERROR("Not implement on cpu");
+}

src/transformers/models/deformable_detr/custom_kernel/cpu/ms_deform_attn_cpu.h

+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#pragma once
+#include <torch/extension.h>
+
+at::Tensor
+ms_deform_attn_cpu_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step);
+
+std::vector<at::Tensor>
+ms_deform_attn_cpu_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step);
+

src/transformers/models/deformable_detr/custom_kernel/cuda/ms_deform_attn_cuda.cu

+/*!
+**************************************************************************************************
+* Deformable DETR
+* Copyright (c) 2020 SenseTime. All Rights Reserved.
+* Licensed under the Apache License, Version 2.0 [see LICENSE for details]
+**************************************************************************************************
+* Modified from https://github.com/chengdazhi/Deformable-Convolution-V2-PyTorch/tree/pytorch_1.0.0
+**************************************************************************************************
+*/
+
+#include <vector>
+#include "cuda/ms_deform_im2col_cuda.cuh"
+
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+
+#pragma once
+#include <torch/extension.h>
+
+
+at::Tensor ms_deform_attn_cuda_forward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const int im2col_step)
+{
+    AT_ASSERTM(value.is_contiguous(), "value tensor has to be contiguous");
+    AT_ASSERTM(spatial_shapes.is_contiguous(), "spatial_shapes tensor has to be contiguous");
+    AT_ASSERTM(level_start_index.is_contiguous(), "level_start_index tensor has to be contiguous");
+    AT_ASSERTM(sampling_loc.is_contiguous(), "sampling_loc tensor has to be contiguous");
+    AT_ASSERTM(attn_weight.is_contiguous(), "attn_weight tensor has to be contiguous");
+
+    AT_ASSERTM(value.type().is_cuda(), "value must be a CUDA tensor");
+    AT_ASSERTM(spatial_shapes.type().is_cuda(), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(level_start_index.type().is_cuda(), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(sampling_loc.type().is_cuda(), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(attn_weight.type().is_cuda(), "attn_weight must be a CUDA tensor");
+
+    const int batch = value.size(0);
+    const int spatial_size = value.size(1);
+    const int num_heads = value.size(2);
+    const int channels = value.size(3);
+
+    const int num_levels = spatial_shapes.size(0);
+
+    const int num_query = sampling_loc.size(1);
+    const int num_point = sampling_loc.size(4);
+
+    const int im2col_step_ = std::min(batch, im2col_step);
+
+    AT_ASSERTM(batch % im2col_step_ == 0, "batch(%d) must divide im2col_step(%d)", batch, im2col_step_);
+    
+    auto output = at::zeros({batch, num_query, num_heads, channels}, value.options());
+
+    const int batch_n = im2col_step_;
+    auto output_n = output.view({batch/im2col_step_, batch_n, num_query, num_heads, channels});
+    auto per_value_size = spatial_size * num_heads * channels;
+    auto per_sample_loc_size = num_query * num_heads * num_levels * num_point * 2;
+    auto per_attn_weight_size = num_query * num_heads * num_levels * num_point;
+    for (int n = 0; n < batch/im2col_step_; ++n)
+    {
+        auto columns = output_n.select(0, n);
+        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_forward_cuda", ([&] {
+            ms_deformable_im2col_cuda(at::cuda::getCurrentCUDAStream(),
+                value.data<scalar_t>() + n * im2col_step_ * per_value_size,
+                spatial_shapes.data<int64_t>(),
+                level_start_index.data<int64_t>(),
+                sampling_loc.data<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                attn_weight.data<scalar_t>() + n * im2col_step_ * per_attn_weight_size,
+                batch_n, spatial_size, num_heads, channels, num_levels, num_query, num_point,
+                columns.data<scalar_t>());
+
+        }));
+    }
+
+    output = output.view({batch, num_query, num_heads*channels});
+
+    return output;
+}
+
+
+std::vector<at::Tensor> ms_deform_attn_cuda_backward(
+    const at::Tensor &value, 
+    const at::Tensor &spatial_shapes,
+    const at::Tensor &level_start_index,
+    const at::Tensor &sampling_loc,
+    const at::Tensor &attn_weight,
+    const at::Tensor &grad_output,
+    const int im2col_step)
+{
+
+    AT_ASSERTM(value.is_contiguous(), "value tensor has to be contiguous");
+    AT_ASSERTM(spatial_shapes.is_contiguous(), "spatial_shapes tensor has to be contiguous");
+    AT_ASSERTM(level_start_index.is_contiguous(), "level_start_index tensor has to be contiguous");
+    AT_ASSERTM(sampling_loc.is_contiguous(), "sampling_loc tensor has to be contiguous");
+    AT_ASSERTM(attn_weight.is_contiguous(), "attn_weight tensor has to be contiguous");
+    AT_ASSERTM(grad_output.is_contiguous(), "grad_output tensor has to be contiguous");
+
+    AT_ASSERTM(value.type().is_cuda(), "value must be a CUDA tensor");
+    AT_ASSERTM(spatial_shapes.type().is_cuda(), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(level_start_index.type().is_cuda(), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(sampling_loc.type().is_cuda(), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(attn_weight.type().is_cuda(), "attn_weight must be a CUDA tensor");
+    AT_ASSERTM(grad_output.type().is_cuda(), "grad_output must be a CUDA tensor");
+
+    const int batch = value.size(0);
+    const int spatial_size = value.size(1);
+    const int num_heads = value.size(2);
+    const int channels = value.size(3);
+
+    const int num_levels = spatial_shapes.size(0);
+
+    const int num_query = sampling_loc.size(1);
+    const int num_point = sampling_loc.size(4);
+
+    const int im2col_step_ = std::min(batch, im2col_step);
+
+    AT_ASSERTM(batch % im2col_step_ == 0, "batch(%d) must divide im2col_step(%d)", batch, im2col_step_);
+
+    auto grad_value = at::zeros_like(value);
+    auto grad_sampling_loc = at::zeros_like(sampling_loc);
+    auto grad_attn_weight = at::zeros_like(attn_weight);
+
+    const int batch_n = im2col_step_;
+    auto per_value_size = spatial_size * num_heads * channels;
+    auto per_sample_loc_size = num_query * num_heads * num_levels * num_point * 2;
+    auto per_attn_weight_size = num_query * num_heads * num_levels * num_point;
+    auto grad_output_n = grad_output.view({batch/im2col_step_, batch_n, num_query, num_heads, channels});
+    
+    for (int n = 0; n < batch/im2col_step_; ++n)
+    {
+        auto grad_output_g = grad_output_n.select(0, n);
+        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_backward_cuda", ([&] {
+            ms_deformable_col2im_cuda(at::cuda::getCurrentCUDAStream(),
+                                    grad_output_g.data<scalar_t>(),
+                                    value.data<scalar_t>() + n * im2col_step_ * per_value_size,
+                                    spatial_shapes.data<int64_t>(),
+                                    level_start_index.data<int64_t>(),
+                                    sampling_loc.data<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                                    attn_weight.data<scalar_t>() + n * im2col_step_ * per_attn_weight_size,
+                                    batch_n, spatial_size, num_heads, channels, num_levels, num_query, num_point,
+                                    grad_value.data<scalar_t>() +  n * im2col_step_ * per_value_size,
+                                    grad_sampling_loc.data<scalar_t>() + n * im2col_step_ * per_sample_loc_size,
+                                    grad_attn_weight.data<scalar_t>() + n * im2col_step_ * per_attn_weight_size);
+
+        }));
+    }
+
+    return {
+        grad_value, grad_sampling_loc, grad_attn_weight
+    };
+}