Add ViTDet (#25524)

* First draft

* Fix READMEs

* Update return_dict

* Add more tests

* Fix docstrings

* Address comments

* Address more comments

* Address more comments

* Address more comments, fix test

* Fix test

README.md

@@ -486,6 +486,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

README_es.md

@@ -463,6 +463,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

README_hd.md

@@ -435,6 +435,7 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (गूगल एआई से) कागज के साथ [एक इमेज इज़ वर्थ 16x16 वर्ड्स: ट्रांसफॉर्मर्स फॉर इमेज रिकॉग्निशन एट स्केल](https://arxiv.org/abs/2010.11929) एलेक्सी डोसोवित्स्की, लुकास बेयर, अलेक्जेंडर कोलेसनिकोव, डिर्क वीसेनबोर्न, शियाओहुआ झाई, थॉमस अनटरथिनर, मुस्तफा देहघानी, मैथियास मिंडरर, जॉर्ज हेगोल्ड, सिल्वेन गेली, जैकब उस्ज़कोरेइट द्वारा हॉल्सबी द्वारा पोस्ट किया गया।
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP से) साथ वाला पेपर [VisualBERT: A Simple and Performant Baseline for Vision and Language](https:/ /arxiv.org/pdf/1908.03557) लियुनियन हेरोल्ड ली, मार्क यात्स्कर, दा यिन, चो-जुई हसीह, काई-वेई चांग द्वारा।
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (Meta AI से) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. द्वाराअनुसंधान पत्र [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) के साथ जारी किया गया
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (मेटा एआई से) साथ में कागज [मास्कड ऑटोएन्कोडर स्केलेबल विजन लर्नर्स हैं](https://arxiv.org/ एब्स/2111.06377) कैमिंग हे, ज़िनेली चेन, सेनिंग ज़ी, यांगहो ली, पिओट्र डॉलर, रॉस गिर्शिक द्वारा।
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (मेटा एआई से) साथ में कागज [लेबल-कुशल सीखने के लिए मास्क्ड स्याम देश के नेटवर्क](https://arxiv. org/abs/2204.07141) महमूद असरान, मथिल्डे कैरन, ईशान मिश्रा, पियोट्र बोजानोवस्की, फ्लोरियन बोर्डेस, पास्कल विंसेंट, आर्मंड जौलिन, माइकल रब्बत, निकोलस बल्लास द्वारा।
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

README_ja.md

@@ -497,6 +497,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (Google AI から) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby から公開された研究論文: [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929)
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP から) Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang から公開された研究論文: [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557)
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (Google AI から) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby から公開された研究論文: [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929)
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (Meta AI から) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. から公開された研究論文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (Meta AI から) Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick から公開された研究論文: [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377)
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (Meta AI から) Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas から公開された研究論文: [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141)
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

README_ko.md

@@ -412,6 +412,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (Google AI 에서) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 의 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 논문과 함께 발표했습니다.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP 에서) Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 의 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 논문과 함께 발표했습니다.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (Google AI 에서) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 의 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 논문과 함께 발표했습니다.
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (Meta AI 에서 제공)은 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.의 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)논문과 함께 발표했습니다.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (Meta AI 에서) Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 의 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 논문과 함께 발표했습니다.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (Meta AI 에서) Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 의 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) 논문과 함께 발표했습니다.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

README_zh-hans.md

@@ -436,6 +436,7 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (来自 UCLA NLP) 伴随论文 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 由 Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 发布。
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (来自 Meta AI) 伴随论文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) 由 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He 发布。
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (来自 Meta AI) 伴随论文 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 由 Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 发布。
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (来自 Meta AI) 伴随论文 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 发布.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (来自 Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) 由 Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

README_zh-hant.md

@@ -448,6 +448,7 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/main/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.

docs/source/en/_toctree.yml

@@ -558,6 +558,8 @@
         title: Vision Transformer (ViT)
       - local: model_doc/vit_hybrid
         title: ViT Hybrid
+      - local: model_doc/vitdet
+        title: ViTDet
       - local: model_doc/vit_mae
         title: ViTMAE
       - local: model_doc/vit_msn

docs/source/en/index.md

@@ -252,6 +252,7 @@ The documentation is organized into five sections:
 1. **[Vision Transformer (ViT)](model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
 1. **[ViViT](model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
@@ -471,6 +472,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          VisualBERT           |       ✅        |         ❌         |      ❌      |
 |              ViT              |       ✅        |         ✅         |      ✅      |
 |          ViT Hybrid           |       ✅        |         ❌         |      ❌      |
+|            VitDet             |       ✅        |         ❌         |      ❌      |
 |            ViTMAE             |       ✅        |         ✅         |      ❌      |
 |            ViTMSN             |       ✅        |         ❌         |      ❌      |
 |             ViViT             |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/vitdet.md

@@ -0,0 +1,39 @@
+<!--Copyright 2023 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.
+-->
+
+# ViTDet
+
+## Overview
+
+The ViTDet model was proposed in [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
+VitDet leverages the plain [Vision Transformer](vit) for the task of object detection.
+
+The abstract from the paper is the following:
+
+*We explore the plain, non-hierarchical Vision Transformer (ViT) as a backbone network for object detection. This design enables the original ViT architecture to be fine-tuned for object detection without needing to redesign a hierarchical backbone for pre-training. With minimal adaptations for fine-tuning, our plain-backbone detector can achieve competitive results. Surprisingly, we observe: (i) it is sufficient to build a simple feature pyramid from a single-scale feature map (without the common FPN design) and (ii) it is sufficient to use window attention (without shifting) aided with very few cross-window propagation blocks. With plain ViT backbones pre-trained as Masked Autoencoders (MAE), our detector, named ViTDet, can compete with the previous leading methods that were all based on hierarchical backbones, reaching up to 61.3 AP_box on the COCO dataset using only ImageNet-1K pre-training. We hope our study will draw attention to research on plain-backbone detectors.*
+
+Tips:
+
+- For the moment, only the backbone is available.
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/facebookresearch/detectron2/tree/main/projects/ViTDet).
+
+
+## VitDetConfig
+
+[[autodoc]] VitDetConfig
+
+## VitDetModel
+
+[[autodoc]] VitDetModel
+    - forward

src/transformers/__init__.py

@@ -586,6 +586,7 @@ _import_structure = {
     "models.vit_hybrid": ["VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTHybridConfig"],
     "models.vit_mae": ["VIT_MAE_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTMAEConfig"],
     "models.vit_msn": ["VIT_MSN_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTMSNConfig"],
+    "models.vitdet": ["VITDET_PRETRAINED_CONFIG_ARCHIVE_MAP", "VitDetConfig"],
     "models.vivit": [
         "VIVIT_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "VivitConfig",
@@ -2925,6 +2926,14 @@ else:
             "ViTMSNPreTrainedModel",
         ]
     )
+    _import_structure["models.vitdet"].extend(
+        [
+            "VITDET_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "VitDetBackbone",
+            "VitDetModel",
+            "VitDetPreTrainedModel",
+        ]
+    )
     _import_structure["models.vivit"].extend(
         [
             "VIVIT_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -4632,6 +4641,7 @@ if TYPE_CHECKING:
     from .models.vit_hybrid import VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTHybridConfig
     from .models.vit_mae import VIT_MAE_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTMAEConfig
     from .models.vit_msn import VIT_MSN_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTMSNConfig
+    from .models.vitdet import VITDET_PRETRAINED_CONFIG_ARCHIVE_MAP, VitDetConfig
     from .models.vivit import VIVIT_PRETRAINED_CONFIG_ARCHIVE_MAP, VivitConfig
     from .models.wav2vec2 import (
         WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -6577,6 +6587,12 @@ if TYPE_CHECKING:
             ViTMSNModel,
             ViTMSNPreTrainedModel,
         )
+        from .models.vitdet import (
+            VITDET_PRETRAINED_MODEL_ARCHIVE_LIST,
+            VitDetBackbone,
+            VitDetModel,
+            VitDetPreTrainedModel,
+        )
         from .models.vivit import (
             VIVIT_PRETRAINED_MODEL_ARCHIVE_LIST,
             VivitForVideoClassification,

src/transformers/models/__init__.py

@@ -210,6 +210,7 @@ from . import (
     vit_hybrid,
     vit_mae,
     vit_msn,
+    vitdet,
     vivit,
     wav2vec2,
     wav2vec2_conformer,

src/transformers/models/auto/configuration_auto.py

@@ -218,6 +218,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("vit_hybrid", "ViTHybridConfig"),
         ("vit_mae", "ViTMAEConfig"),
         ("vit_msn", "ViTMSNConfig"),
+        ("vitdet", "VitDetConfig"),
         ("vivit", "VivitConfig"),
         ("wav2vec2", "Wav2Vec2Config"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerConfig"),
@@ -408,6 +409,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("vit_hybrid", "VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vit_mae", "VIT_MAE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vit_msn", "VIT_MSN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vitdet", "VITDET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vivit", "VIVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("wav2vec2", "WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("wav2vec2-conformer", "WAV2VEC2_CONFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -640,6 +642,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("vit_hybrid", "ViT Hybrid"),
         ("vit_mae", "ViTMAE"),
         ("vit_msn", "ViTMSN"),
+        ("vitdet", "VitDet"),
         ("vivit", "ViViT"),
         ("wav2vec2", "Wav2Vec2"),
         ("wav2vec2-conformer", "Wav2Vec2-Conformer"),

src/transformers/models/auto/modeling_auto.py

@@ -204,6 +204,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("vit_hybrid", "ViTHybridModel"),
         ("vit_mae", "ViTMAEModel"),
         ("vit_msn", "ViTMSNModel"),
+        ("vitdet", "VitDetModel"),
         ("vivit", "VivitModel"),
         ("wav2vec2", "Wav2Vec2Model"),
         ("wav2vec2-conformer", "Wav2Vec2ConformerModel"),
@@ -1061,6 +1062,7 @@ MODEL_FOR_BACKBONE_MAPPING_NAMES = OrderedDict(
         ("resnet", "ResNetBackbone"),
         ("swin", "SwinBackbone"),
         ("timm_backbone", "TimmBackbone"),
+        ("vitdet", "VitDetBackbone"),
     ]
 )
 

src/transformers/models/vitdet/__init__.py

@@ -0,0 +1,57 @@
+# Copyright 2023 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,
+)
+
+
+_import_structure = {"configuration_vitdet": ["VITDET_PRETRAINED_CONFIG_ARCHIVE_MAP", "VitDetConfig"]}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_vitdet"] = [
+        "VITDET_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "VitDetModel",
+        "VitDetPreTrainedModel",
+        "VitDetBackbone",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_vitdet import VITDET_PRETRAINED_CONFIG_ARCHIVE_MAP, VitDetConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_vitdet import (
+            VITDET_PRETRAINED_MODEL_ARCHIVE_LIST,
+            VitDetBackbone,
+            VitDetModel,
+            VitDetPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/vitdet/configuration_vitdet.py

@@ -0,0 +1,155 @@
+# coding=utf-8
+# Copyright 2023 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.
+""" VitDet 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__)
+
+VITDET_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/vit-det-base": "https://huggingface.co/facebook/vit-det-base/resolve/main/config.json",
+}
+
+
+class VitDetConfig(BackboneConfigMixin, PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`VitDetModel`]. It is used to instantiate an
+    VitDet 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 VitDet
+    [google/vitdet-base-patch16-224](https://huggingface.co/google/vitdet-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        mlp_ratio (`int`, *optional*, defaults to 4):
+            Ratio of mlp hidden dim to embedding dim.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-6):
+            The epsilon used by the layer normalization layers.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        pretrain_image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image during pretraining.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries, keys and values.
+        drop_path_rate (`float`, *optional*, defaults to 0.0):
+            Stochastic depth rate.
+        window_block_indices (`List[int]`, *optional*):
+            List of indices of blocks that should have window attention instead of regular global self-attention.
+        residual_block_indices (`List[int]`, *optional*):
+            List of indices of blocks that should have an extra residual block after the MLP.
+        use_absolute_position_embeddings (`bool`, *optional*, defaults to `True`):
+            Whether to add absolute position embeddings to the patch embeddings.
+        use_relative_position_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to add relative position embeddings to the attention maps.
+        window_size (`int`, *optional*, defaults to 0):
+            The size of the attention window.
+        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.
+        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.
+
+    Example:
+
+    ```python
+    >>> from transformers import VitDetConfig, VitDetModel
+
+    >>> # Initializing a VitDet configuration
+    >>> configuration = VitDetConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = VitDetModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "vitdet"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        mlp_ratio=4,
+        hidden_act="gelu",
+        dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-6,
+        image_size=224,
+        pretrain_image_size=224,
+        patch_size=16,
+        num_channels=3,
+        qkv_bias=True,
+        drop_path_rate=0.0,
+        window_block_indices=[],
+        residual_block_indices=[],
+        use_absolute_position_embeddings=True,
+        use_relative_position_embeddings=False,
+        window_size=0,
+        out_features=None,
+        out_indices=None,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.mlp_ratio = mlp_ratio
+        self.hidden_act = hidden_act
+        self.dropout_prob = dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.image_size = image_size
+        self.pretrain_image_size = pretrain_image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.qkv_bias = qkv_bias
+        self.drop_path_rate = drop_path_rate
+        self.window_block_indices = window_block_indices
+        self.residual_block_indices = residual_block_indices
+        self.use_absolute_position_embeddings = use_absolute_position_embeddings
+        self.use_relative_position_embeddings = use_relative_position_embeddings
+        self.window_size = window_size
+
+        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, self.num_hidden_layers + 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/vitdet/modeling_vitdet.py

@@ -0,0 +1,875 @@
+# coding=utf-8
+# Copyright 2023 Meta AI 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 ViTDet backbone."""
+
+
+import collections.abc
+import math
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...activations import ACT2FN
+from ...modeling_outputs import BackboneOutput, BaseModelOutput
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from ...utils.backbone_utils import BackboneMixin
+from .configuration_vitdet import VitDetConfig
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "VitDetConfig"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "facebook/vit-det-base"
+_EXPECTED_OUTPUT_SHAPE = [1, 197, 768]
+
+
+VITDET_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "facebook/vit-det-base",
+    # See all ViTDet models at https://huggingface.co/models?filter=vitdet
+]
+
+
+class VitDetEmbeddings(nn.Module):
+    """
+    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
+    `hidden_states` (patch embeddings) to be consumed by a Transformer.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        image_size, patch_size = config.pretrain_image_size, config.patch_size
+        num_channels, hidden_size = config.num_channels, config.hidden_size
+
+        image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size)
+        patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.num_patches = num_patches
+
+        if config.use_absolute_position_embeddings:
+            # Initialize absolute positional embedding with pretrain image size.
+            num_positions = num_patches + 1
+            self.position_embeddings = nn.Parameter(torch.zeros(1, num_positions, config.hidden_size))
+        else:
+            self.position_embeddings = None
+
+        self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size)
+
+    def get_absolute_positions(self, abs_pos_embeddings, has_cls_token, height, width):
+        """
+        Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token dimension for the
+        original embeddings.
+
+        Args:
+            abs_pos_embeddings (`torch.Tensor`):
+                Absolute positional embeddings with (1, num_position, num_channels).
+            has_cls_token (`bool`):
+                If true, has 1 embedding in abs_pos_embeddings for cls token.
+            height (`int`):
+                Height of input image tokens.
+            width (`int`):
+                Width of input image tokens.
+
+        Returns:
+            Absolute positional embeddings after processing with shape (1, height, width, num_channels)
+        """
+        if has_cls_token:
+            abs_pos_embeddings = abs_pos_embeddings[:, 1:]
+        num_position = abs_pos_embeddings.shape[1]
+        size = int(math.sqrt(num_position))
+        if size * size != num_position:
+            raise ValueError("Absolute position embeddings must be a square number.")
+
+        if size != height or size != width:
+            new_abs_pos_embeddings = nn.functional.interpolate(
+                abs_pos_embeddings.reshape(1, size, size, -1).permute(0, 3, 1, 2),
+                size=(height, width),
+                mode="bicubic",
+                align_corners=False,
+            )
+
+            return new_abs_pos_embeddings.permute(0, 2, 3, 1)
+        else:
+            return abs_pos_embeddings.reshape(1, height, width, -1)
+
+    def forward(self, pixel_values: torch.Tensor) -> torch.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."
+                f" Expected {self.num_channels} but got {num_channels}."
+            )
+        embeddings = self.projection(pixel_values)
+
+        if self.position_embeddings is not None:
+            # (batch_size, num_channels, height, width) -> (batch_size, height, width, num_channels)
+            embeddings = embeddings.permute(0, 2, 3, 1)
+            # add position embeddings
+            embeddings = embeddings + self.get_absolute_positions(
+                self.position_embeddings, True, embeddings.shape[1], embeddings.shape[2]
+            )
+            # (batch_size, height, width, num_channels) -> (batch_size, num_channels, height, width)
+            embeddings = embeddings.permute(0, 3, 1, 2)
+
+        return embeddings
+
+
+def get_rel_pos(q_size, k_size, rel_pos):
+    """
+    Get relative positional embeddings according to the relative positions of query and key sizes.
+
+    Args:
+        q_size (`int`):
+            Size of query q.
+        k_size (`int`):
+            Size of key k.
+        rel_pos (`torch.Tensor`):
+            Relative position embeddings (num_embeddings, num_channels).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel position embeddings.
+        rel_pos_resized = nn.functional.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_relative_positions(attn, queries, rel_pos_h, rel_pos_w, q_size, k_size):
+    """
+    Calculate decomposed Relative Positional Embeddings as introduced in
+    [MViT2](https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py).
+
+    Args:
+        attn (`torch.Tensor`):
+            Attention map.
+        queries (`torch.Tensor`):
+            Query q in the attention layer with shape (batch_size, queries_height * queries_width, num_channels).
+        rel_pos_h (`torch.Tensor`):
+            Relative position embeddings (Lh, num_channels) for height axis.
+        rel_pos_w (`torch.Tensor`):
+            Relative position embeddings (Lw, num_channels) for width axis.
+        q_size (`Tuple[int]`):
+            Spatial sequence size of query q with (queries_height, queries_width).
+        k_size (`Tuple[int]`]):
+            Spatial sequence size of key k with (keys_height, keys_width).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    queries_height, queries_width = q_size
+    keys_height, keys_width = k_size
+    relative_height = get_rel_pos(queries_height, keys_height, rel_pos_h)
+    relative_width = get_rel_pos(queries_width, keys_width, rel_pos_w)
+
+    batch_size, _, dim = queries.shape
+    r_q = queries.reshape(batch_size, queries_height, queries_width, dim)
+    relative_height = torch.einsum("bhwc,hkc->bhwk", r_q, relative_height)
+    relative_weight = torch.einsum("bhwc,wkc->bhwk", r_q, relative_width)
+
+    attn = (
+        attn.view(batch_size, queries_height, queries_width, keys_height, keys_width)
+        + relative_height[:, :, :, :, None]
+        + relative_weight[:, :, :, None, :]
+    ).view(batch_size, queries_height * queries_width, keys_height * keys_width)
+
+    return attn
+
+
+class VitDetAttention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(self, config, input_size=None):
+        """
+        Args:
+            config (`VitDetConfig`):
+                Model configuration.
+            input_size (`Tuple[int]`, *optional*):
+                Input resolution, only required in case relative position embeddings are added.
+        """
+        super().__init__()
+
+        dim = config.hidden_size
+        num_heads = config.num_attention_heads
+
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=config.qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_relative_position_embeddings = config.use_relative_position_embeddings
+        if self.use_relative_position_embeddings:
+            # initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, hidden_state, output_attentions=False):
+        batch_size, height, width, _ = hidden_state.shape
+        # qkv with shape (3, batch_size, num_heads, height * width, num_channels)
+        qkv = self.qkv(hidden_state).reshape(batch_size, height * width, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        # queries, keys and values have shape (batch_size * num_heads, height * width, num_channels)
+        queries, keys, values = qkv.reshape(3, batch_size * self.num_heads, height * width, -1).unbind(0)
+
+        attention_scores = (queries * self.scale) @ keys.transpose(-2, -1)
+
+        if self.use_relative_position_embeddings:
+            attention_scores = add_decomposed_relative_positions(
+                attention_scores, queries, self.rel_pos_h, self.rel_pos_w, (height, width), (height, width)
+            )
+
+        attention_probs = attention_scores.softmax(dim=-1)
+
+        hidden_state = attention_probs @ values
+        hidden_state = hidden_state.view(batch_size, self.num_heads, height, width, -1)
+        hidden_state = hidden_state.permute(0, 2, 3, 1, 4)
+        hidden_state = hidden_state.reshape(batch_size, height, width, -1)
+        hidden_state = self.proj(hidden_state)
+
+        if output_attentions:
+            attention_probs = attention_probs.reshape(
+                batch_size, self.num_heads, attention_probs.shape[-2], attention_probs.shape[-1]
+            )
+            outputs = (hidden_state, attention_probs)
+        else:
+            outputs = (hidden_state,)
+
+        return outputs
+
+
+# Copied from transformers.models.beit.modeling_beit.drop_path
+def drop_path(input: torch.Tensor, drop_prob: float = 0.0, training: bool = False) -> torch.Tensor:
+    """
+    Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    Comment by Ross Wightman: This is the same as the DropConnect impl I created for EfficientNet, etc networks,
+    however, the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for changing the
+    layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use 'survival rate' as the
+    argument.
+    """
+    if drop_prob == 0.0 or not training:
+        return input
+    keep_prob = 1 - drop_prob
+    shape = (input.shape[0],) + (1,) * (input.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=input.dtype, device=input.device)
+    random_tensor.floor_()  # binarize
+    output = input.div(keep_prob) * random_tensor
+    return output
+
+
+# Copied from transformers.models.beit.modeling_beit.BeitDropPath
+class VitDetDropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob: Optional[float] = None) -> None:
+        super().__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return drop_path(hidden_states, self.drop_prob, self.training)
+
+    def extra_repr(self) -> str:
+        return "p={}".format(self.drop_prob)
+
+
+class VitDetLayerNorm(nn.Module):
+    """
+    A LayerNorm variant, popularized by Transformers, that performs point-wise mean and variance normalization over the
+    channel dimension for inputs that have shape (batch_size, channels, height, width).
+    https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119
+    """
+
+    def __init__(self, normalized_shape, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(normalized_shape))
+        self.bias = nn.Parameter(torch.zeros(normalized_shape))
+        self.eps = eps
+        self.normalized_shape = (normalized_shape,)
+
+    def forward(self, x):
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+class VitDetResBottleneckBlock(nn.Module):
+    """
+    The standard bottleneck residual block without the last activation layer. It contains 3 conv layers with kernels
+    1x1, 3x3, 1x1.
+    """
+
+    def __init__(self, config, in_channels, out_channels, bottleneck_channels):
+        """
+        Args:
+            config (`VitDetConfig`):
+                Model configuration.
+            in_channels (`int`):
+                Number of input channels.
+            out_channels (`int`):
+                Number of output channels.
+            bottleneck_channels (`int`):
+                Number of output channels for the 3x3 "bottleneck" conv layers.
+        """
+        super().__init__()
+        self.conv1 = nn.Conv2d(in_channels, bottleneck_channels, 1, bias=False)
+        self.norm1 = VitDetLayerNorm(bottleneck_channels)
+        self.act1 = ACT2FN[config.hidden_act]
+
+        self.conv2 = nn.Conv2d(bottleneck_channels, bottleneck_channels, 3, padding=1, bias=False)
+        self.norm2 = VitDetLayerNorm(bottleneck_channels)
+        self.act2 = ACT2FN[config.hidden_act]
+
+        self.conv3 = nn.Conv2d(bottleneck_channels, out_channels, 1, bias=False)
+        self.norm3 = VitDetLayerNorm(out_channels)
+
+    def forward(self, x):
+        out = x
+        for layer in self.children():
+            out = layer(out)
+
+        out = x + out
+        return out
+
+
+class VitDetMlp(nn.Module):
+    def __init__(self, config, in_features: int, hidden_features: int) -> None:
+        super().__init__()
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = ACT2FN[config.hidden_act]
+        self.fc2 = nn.Linear(hidden_features, in_features)
+        self.drop = nn.Dropout(config.dropout_prob)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+
+        return x
+
+
+def window_partition(hidden_state, window_size):
+    """
+    Partition into non-overlapping windows with padding if needed.
+
+    Args:
+        hidden_state (`torch.Tensor`):
+            Input tokens with [batch_size, height, width, num_channels].
+        window_size (`int`):
+            Window size.
+
+    Returns:
+        `tuple(torch.FloatTensor)` comprising various elements:
+        - windows: windows after partition with [batch_size * num_windows, window_size, window_size, num_channels].
+        - (patch_height, patch_width): padded height and width before partition
+    """
+    batch_size, height, width, num_channels = hidden_state.shape
+
+    pad_height = (window_size - height % window_size) % window_size
+    pad_width = (window_size - width % window_size) % window_size
+    if pad_height > 0 or pad_width > 0:
+        hidden_state = nn.functional.pad(hidden_state, (0, 0, 0, pad_width, 0, pad_height))
+    patch_height, patch_width = height + pad_height, width + pad_width
+
+    hidden_state = hidden_state.view(
+        batch_size, patch_height // window_size, window_size, patch_width // window_size, window_size, num_channels
+    )
+    windows = hidden_state.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, num_channels)
+    return windows, (patch_height, patch_width)
+
+
+def window_unpartition(windows, window_size, pad_height_width, height_width):
+    """
+    Window unpartition into original sequences and removing padding.
+
+    Args:
+        windows (`torch.Tensor`):
+            Input tokens with [batch_size * num_windows, window_size, window_size, num_channels].
+        window_size (`int`):
+            Window size.
+        pad_height_width (`Tuple[int]`):
+            Padded height and width (patch_height, patch_width).
+        height_width (`Tuple[int]`):
+            Original height and width before padding.
+
+    Returns:
+        hidden_state: unpartitioned sequences with [batch_size, height, width, num_channels].
+    """
+    patch_height, patch_width = pad_height_width
+    height, width = height_width
+    batch_size = windows.shape[0] // (patch_height * patch_width // window_size // window_size)
+    hidden_state = windows.view(
+        batch_size, patch_height // window_size, patch_width // window_size, window_size, window_size, -1
+    )
+    hidden_state = hidden_state.permute(0, 1, 3, 2, 4, 5).contiguous().view(batch_size, patch_height, patch_width, -1)
+
+    if patch_height > height or patch_width > width:
+        hidden_state = hidden_state[:, :height, :width, :].contiguous()
+    return hidden_state
+
+
+class VitDetLayer(nn.Module):
+    """This corresponds to the Block class in the original implementation."""
+
+    def __init__(
+        self, config: VitDetConfig, drop_path_rate: float = 0, window_size: int = 0, use_residual_block: bool = False
+    ) -> None:
+        super().__init__()
+
+        dim = config.hidden_size
+        input_size = (config.image_size // config.patch_size, config.image_size // config.patch_size)
+
+        self.norm1 = nn.LayerNorm(dim, eps=config.layer_norm_eps)
+        self.attention = VitDetAttention(
+            config, input_size=input_size if window_size == 0 else (window_size, window_size)
+        )
+
+        self.drop_path = VitDetDropPath(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
+        self.norm2 = nn.LayerNorm(dim, eps=config.layer_norm_eps)
+        self.mlp = VitDetMlp(config=config, in_features=dim, hidden_features=int(dim * config.mlp_ratio))
+
+        self.window_size = window_size
+
+        self.use_residual_block = use_residual_block
+        if self.use_residual_block:
+            # Use a residual block with bottleneck channel as dim // 2
+            self.residual = VitDetResBottleneckBlock(
+                config=config,
+                in_channels=dim,
+                out_channels=dim,
+                bottleneck_channels=dim // 2,
+            )
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        hidden_states = hidden_states.permute(0, 2, 3, 1)
+
+        shortcut = hidden_states
+
+        hidden_states = self.norm1(hidden_states)
+
+        # Window partition
+        if self.window_size > 0:
+            height, width = hidden_states.shape[1], hidden_states.shape[2]
+            hidden_states, pad_height_width = window_partition(hidden_states, self.window_size)
+
+        self_attention_outputs = self.attention(
+            hidden_states,
+            output_attentions=output_attentions,
+        )
+        hidden_states = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        # Reverse window partition
+        if self.window_size > 0:
+            hidden_states = window_unpartition(hidden_states, self.window_size, pad_height_width, (height, width))
+
+        # first residual connection
+        hidden_states = shortcut + self.drop_path(hidden_states)
+
+        hidden_states = hidden_states + self.drop_path(self.mlp(self.norm2(hidden_states)))
+
+        hidden_states = hidden_states.permute(0, 3, 1, 2)
+
+        if self.use_residual_block:
+            hidden_states = self.residual(hidden_states)
+
+        outputs = (hidden_states,) + outputs
+
+        return outputs
+
+
+class VitDetEncoder(nn.Module):
+    def __init__(self, config: VitDetConfig) -> None:
+        super().__init__()
+        self.config = config
+        depth = config.num_hidden_layers
+
+        # stochastic depth decay rule
+        drop_path_rate = [x.item() for x in torch.linspace(0, config.drop_path_rate, depth)]
+
+        layers = []
+        for i in range(depth):
+            layers.append(
+                VitDetLayer(
+                    config,
+                    drop_path_rate=drop_path_rate[i],
+                    window_size=config.window_size if i in config.window_block_indices else 0,
+                    use_residual_block=i in config.residual_block_indices,
+                )
+            )
+
+        self.layer = nn.ModuleList(layers)
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ) -> Union[tuple, BaseModelOutput]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    layer_head_mask,
+                )
+            else:
+                layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, all_hidden_states, all_self_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+        )
+
+
+def caffe2_msra_fill(module: nn.Module) -> None:
+    """
+    Initialize `module.weight` using the "MSRAFill" implemented in Caffe2. Also initializes `module.bias` to 0.
+
+    Source: https://detectron2.readthedocs.io/en/latest/_modules/fvcore/nn/weight_init.html.
+
+    Args:
+        module (torch.nn.Module): module to initialize.
+    """
+    nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
+    if module.bias is not None:
+        nn.init.constant_(module.bias, 0)
+
+
+class VitDetPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = VitDetConfig
+    base_model_prefix = "vitdet"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+    _no_split_modules = []
+
+    def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> None:
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Conv2d)):
+            # Upcast the input in `fp32` and cast it back to desired `dtype` to avoid
+            # `trunc_normal_cpu` not implemented in `half` issues
+            module.weight.data = nn.init.trunc_normal_(
+                module.weight.data.to(torch.float32), mean=0.0, std=self.config.initializer_range
+            ).to(module.weight.dtype)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+        elif isinstance(module, VitDetEmbeddings):
+            module.position_embeddings.data = nn.init.trunc_normal_(
+                module.position_embeddings.data.to(torch.float32),
+                mean=0.0,
+                std=self.config.initializer_range,
+            ).to(module.position_embeddings.dtype)
+
+        elif isinstance(module, VitDetAttention) and self.config.use_relative_position_embeddings:
+            module.rel_pos_h.data = nn.init.trunc_normal_(
+                module.rel_pos_h.data.to(torch.float32),
+                mean=0.0,
+                std=self.config.initializer_range,
+            )
+            module.rel_pos_w.data = nn.init.trunc_normal_(
+                module.rel_pos_w.data.to(torch.float32),
+                mean=0.0,
+                std=self.config.initializer_range,
+            )
+
+        elif isinstance(module, VitDetResBottleneckBlock):
+            for layer in [module.conv1, module.conv2, module.conv3]:
+                caffe2_msra_fill(layer)
+            for layer in [module.norm1, module.norm2]:
+                layer.weight.data.fill_(1.0)
+                layer.bias.data.zero_()
+            # zero init last norm layer.
+            module.norm3.weight.data.zero_()
+            module.norm3.bias.data.zero_()
+
+    def _set_gradient_checkpointing(self, module: VitDetEncoder, value: bool = False) -> None:
+        if isinstance(module, VitDetEncoder):
+            module.gradient_checkpointing = value
+
+
+VITDET_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 ([`VitDetConfig`]): 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.
+"""
+
+VITDET_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 [`ViTImageProcessor.__call__`]
+            for details.
+
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        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(
+    "The bare VitDet Transformer model outputting raw hidden-states without any specific head on top.",
+    VITDET_START_DOCSTRING,
+)
+class VitDetModel(VitDetPreTrainedModel):
+    def __init__(self, config: VitDetConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = VitDetEmbeddings(config)
+        self.encoder = VitDetEncoder(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> VitDetEmbeddings:
+        return self.embeddings.projection
+
+    def _prune_heads(self, heads_to_prune: Dict[int, List[int]]) -> None:
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(VITDET_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+        modality="vision",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(pixel_values)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+
+        if not return_dict:
+            return (sequence_output,) + encoder_outputs[1:]
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    ViTDet backbone, to be used with frameworks like Mask R-CNN.
+    """,
+    VITDET_START_DOCSTRING,
+)
+class VitDetBackbone(VitDetPreTrainedModel, BackboneMixin):
+    def __init__(self, config):
+        super().__init__(config)
+        super()._init_backbone(config)
+
+        self.embeddings = VitDetEmbeddings(config)
+        self.encoder = VitDetEncoder(config)
+        self.num_features = [config.hidden_size for _ in range(config.num_hidden_layers + 1)]
+
+        # initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> VitDetEmbeddings:
+        return self.embeddings.projection
+
+    @add_start_docstrings_to_model_forward(VITDET_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BackboneOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> BackboneOutput:
+        """
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import AutoImageProcessor, AutoBackbone
+        >>> import torch
+        >>> from PIL import Image
+        >>> import requests
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> processor = AutoImageProcessor.from_pretrained("facebook/convnext-tiny-224")
+        >>> model = AutoBackbone.from_pretrained("facebook/convnext-tiny-224")
+
+        >>> inputs = processor(image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        ```"""
+        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
+        )
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+
+        embedding_output = self.embeddings(pixel_values)
+
+        outputs = self.encoder(
+            embedding_output,
+            output_hidden_states=True,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs.hidden_states if return_dict else outputs[1]
+
+        feature_maps = ()
+        for stage, hidden_state in zip(self.stage_names, hidden_states):
+            if stage in self.out_features:
+                feature_maps += (hidden_state,)
+
+        if not return_dict:
+            if output_hidden_states:
+                output = (feature_maps,) + outputs[1:]
+            else:
+                output = (feature_maps,) + outputs[2:]
+            return output
+
+        return BackboneOutput(
+            feature_maps=feature_maps,
+            hidden_states=outputs.hidden_states if output_hidden_states else None,
+            attentions=outputs.attentions,
+        )

src/transformers/utils/dummy_pt_objects.py

@@ -7898,6 +7898,30 @@ class ViTMSNPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+VITDET_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class VitDetBackbone(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class VitDetModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class VitDetPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 VIVIT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/vitdet/test_modeling_vitdet.py

@@ -0,0 +1,291 @@
+# coding=utf-8
+# Copyright 2023 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.
+""" Testing suite for the PyTorch ViTDet model. """
+
+
+import inspect
+import unittest
+
+from transformers import VitDetConfig
+from transformers.testing_utils import require_torch, torch_device
+from transformers.utils import is_torch_available
+
+from ...test_backbone_common import BackboneTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import VitDetBackbone, VitDetModel
+
+
+class VitDetModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        type_sequence_label_size=10,
+        initializer_range=0.02,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+        self.num_patches_one_direction = self.image_size // self.patch_size
+        self.seq_length = (self.image_size // self.patch_size) ** 2
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+
+        config = self.get_config()
+
+        return config, pixel_values, labels
+
+    def get_config(self):
+        return VitDetConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = VitDetModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(
+            result.last_hidden_state.shape,
+            (self.batch_size, self.hidden_size, self.num_patches_one_direction, self.num_patches_one_direction),
+        )
+
+    def create_and_check_backbone(self, config, pixel_values, labels):
+        model = VitDetBackbone(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+
+        # verify hidden states
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertListEqual(
+            list(result.feature_maps[0].shape),
+            [self.batch_size, self.hidden_size, self.num_patches_one_direction, self.num_patches_one_direction],
+        )
+
+        # verify channels
+        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertListEqual(model.channels, [config.hidden_size])
+
+        # verify backbone works with out_features=None
+        config.out_features = None
+        model = VitDetBackbone(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+
+        # verify feature maps
+        self.parent.assertEqual(len(result.feature_maps), 1)
+        self.parent.assertListEqual(
+            list(result.feature_maps[0].shape),
+            [self.batch_size, self.hidden_size, self.num_patches_one_direction, self.num_patches_one_direction],
+        )
+
+        # verify channels
+        self.parent.assertEqual(len(model.channels), 1)
+        self.parent.assertListEqual(model.channels, [config.hidden_size])
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values, labels = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class VitDetModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as VitDet does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (VitDetModel, VitDetBackbone) if is_torch_available() else ()
+    pipeline_model_mapping = {"feature-extraction": VitDetModel} if is_torch_available() else {}
+
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = VitDetModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=VitDetConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="VitDet does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_backbone(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_backbone(*config_and_inputs)
+
+    def test_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            hidden_states = outputs.hidden_states
+
+            expected_num_stages = self.model_tester.num_hidden_layers
+            self.assertEqual(len(hidden_states), expected_num_stages + 1)
+
+            # VitDet's feature maps are of shape (batch_size, num_channels, height, width)
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [
+                    self.model_tester.num_patches_one_direction,
+                    self.model_tester.num_patches_one_direction,
+                ],
+            )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    # overwrite since VitDet only supports retraining gradients of hidden states
+    def test_retain_grad_hidden_states_attentions(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.output_hidden_states = True
+        config.output_attentions = self.has_attentions
+
+        # no need to test all models as different heads yield the same functionality
+        model_class = self.all_model_classes[0]
+        model = model_class(config)
+        model.to(torch_device)
+
+        inputs = self._prepare_for_class(inputs_dict, model_class)
+
+        outputs = model(**inputs)
+
+        output = outputs[0]
+
+        # Encoder-/Decoder-only models
+        hidden_states = outputs.hidden_states[0]
+        hidden_states.retain_grad()
+
+        output.flatten()[0].backward(retain_graph=True)
+
+        self.assertIsNotNone(hidden_states.grad)
+
+    @unittest.skip(reason="VitDet does not support feedforward chunking")
+    def test_feed_forward_chunking(self):
+        pass
+
+    @unittest.skip(reason="VitDet does not have standalone checkpoints since it used as backbone in other models")
+    def test_model_from_pretrained(self):
+        pass
+
+
+@require_torch
+class VitDetBackboneTest(unittest.TestCase, BackboneTesterMixin):
+    all_model_classes = (VitDetBackbone,) if is_torch_available() else ()
+    config_class = VitDetConfig
+
+    has_attentions = False
+
+    def setUp(self):
+        self.model_tester = VitDetModelTester(self)

tests/test_backbone_common.py

@@ -31,7 +31,8 @@ class BackboneTesterMixin:
         # test default config
         config = config_class()
         self.assertIsNotNone(config)
-        expected_stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, len(config.depths) + 1)]
+        num_stages = len(config.depths) if hasattr(config, "depths") else config.num_hidden_layers
+        expected_stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, num_stages + 1)]
         self.assertEqual(config.stage_names, expected_stage_names)
         self.assertTrue(set(config.out_features).issubset(set(config.stage_names)))
 

utils/check_repo.py

@@ -965,6 +965,7 @@ SHOULD_HAVE_THEIR_OWN_PAGE = [
     "SwinBackbone",
     "TimmBackbone",
     "TimmBackboneConfig",
+    "VitDetBackbone",
 ]