Add DINOv2 (#24016)

* First draft * More improvements * Convert patch embedding layer * Convert all weights * Make conversion work * Improve conversion script * Fix style * Make all tests pass * Add image processor to auto mapping * Add swiglu ffn * Add image processor to conversion script * Fix conversion of giant model * Fix documentation * Fix style * Fix tests * Address comments * Address more comments * Remove unused arguments * Remove more arguments * Rename parameters * Include mask token * Address comments * Add docstring * Transfer checkpoints * Empty commit

Add DINOv2 (#24016)
* First draft * More improvements * Convert patch embedding layer * Convert all weights * Make conversion work * Improve conversion script * Fix style * Make all tests pass * Add image processor to auto mapping * Add swiglu ffn * Add image processor to conversion script * Fix conversion of giant model * Fix documentation * Fix style * Fix tests * Address comments * Address more comments * Remove unused arguments * Remove more arguments * Rename parameters * Include mask token * Address comments * Add docstring * Transfer checkpoints * Empty commit
3ec10e6c · NielsRogge · GitHub · 57da42ad · 3ec10e6c · 3ec10e6c
Unverified Commit 3ec10e6c authored Jul 18, 2023 by NielsRogge Committed by GitHub Jul 18, 2023
20 changed files
--- a/README.md
+++ b/README.md
@@ -337,6 +337,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 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.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.

--- a/README_es.md
+++ b/README_es.md
@@ -314,6 +314,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 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.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.

--- a/README_hd.md
+++ b/README_hd.md
@@ -286,6 +286,7 @@ conda install -c huggingface transformers
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (फेसबुक से) साथ में कागज [ट्रांसफॉर्मर्स के साथ एंड-टू-एंड ऑब्जेक्ट डिटेक्शन](https://arxiv. org/abs/2005.12872) निकोलस कैरियन, फ़्रांसिस्को मस्सा, गेब्रियल सिनेव, निकोलस उसुनियर, अलेक्जेंडर किरिलोव, सर्गेई ज़ागोरुयको द्वारा।
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (माइक्रोसॉफ्ट रिसर्च से) कागज के साथ [DialoGPT: बड़े पैमाने पर जनरेटिव प्री-ट्रेनिंग फॉर कन्वर्सेशनल रिस्पांस जेनरेशन](https ://arxiv.org/abs/1911.00536) यिज़े झांग, सिकी सन, मिशेल गैली, येन-चुन चेन, क्रिस ब्रोकेट, जियांग गाओ, जियानफेंग गाओ, जिंगजिंग लियू, बिल डोलन द्वारा।
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (Meta AI से) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski. द्वाराअनुसंधान पत्र [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) के साथ जारी किया गया
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (हगिंगफेस से), साथ में कागज [डिस्टिलबर्ट, बीईआरटी का डिस्टिल्ड वर्जन: छोटा, तेज, सस्ता और हल्का] (https://arxiv.org/abs/1910.01108) विक्टर सनह, लिसांड्रे डेब्यू और थॉमस वुल्फ द्वारा पोस्ट किया गया। यही तरीका GPT-2 को [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERta से [DistilRoBERta](https://github.com) पर कंप्रेस करने के लिए भी लागू किया जाता है। / हगिंगफेस/ट्रांसफॉर्मर्स/ट्री/मेन/उदाहरण/डिस्टिलेशन), बहुभाषी BERT से [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) और डिस्टिलबर्ट का जर्मन संस्करण।
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (माइक्रोसॉफ्ट रिसर्च से) साथ में पेपर [DiT: सेल्फ सुपरवाइज्ड प्री-ट्रेनिंग फॉर डॉक्यूमेंट इमेज ट्रांसफॉर्मर](https://arxiv.org/abs/2203.02378) जुनलॉन्ग ली, यिहेंग जू, टेंगचाओ लव, लेई कुई, चा झांग द्वारा फुरु वेई द्वारा पोस्ट किया गया।
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (NAVER से) साथ में कागज [OCR-मुक्त डॉक्यूमेंट अंडरस्टैंडिंग ट्रांसफॉर्मर](https://arxiv.org/abs /2111.15664) गीवूक किम, टीकग्यू होंग, मूनबिन यिम, जियोंग्योन नाम, जिनयॉन्ग पार्क, जिनयॉन्ग यिम, वोनसेओक ह्वांग, सांगडू यूं, डोंगयून हान, सेउंग्युन पार्क द्वारा।

--- a/README_ja.md
+++ b/README_ja.md
@@ -348,6 +348,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (Facebook から) Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko から公開された研究論文: [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872)
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (Microsoft Research から) Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan から公開された研究論文: [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536)
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (SHI Labs から) Ali Hassani and Humphrey Shi から公開された研究論文: [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001)
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (Meta AI から) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski. から公開された研究論文 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (HuggingFace から), Victor Sanh, Lysandre Debut and Thomas Wolf. 同じ手法で GPT2, RoBERTa と Multilingual BERT の圧縮を行いました.圧縮されたモデルはそれぞれ [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation)、[DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation)、[DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) と名付けられました. 公開された研究論文: [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108)
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (Microsoft Research から) Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei から公開された研究論文: [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378)
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (NAVER から), Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park から公開された研究論文: [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664)

--- a/README_ko.md
+++ b/README_ko.md
@@ -263,6 +263,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (Facebook 에서) Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko 의 [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) 논문과 함께 발표했습니다.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (Microsoft Research 에서) Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan 의 [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) 논문과 함께 발표했습니다.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (SHI Labs 에서) Ali Hassani and Humphrey Shi 의 [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) 논문과 함께 발표했습니다.
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (Meta AI 에서 제공)은 Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.의 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)논문과 함께 발표했습니다.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (HuggingFace 에서) Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) and a German version of DistilBERT 의 [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) 논문과 함께 발표했습니다.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (Microsoft Research 에서) Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei 의 [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) 논문과 함께 발표했습니다.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (NAVER 에서) Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park 의 [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) 논문과 함께 발표했습니다.

--- a/README_zh-hans.md
+++ b/README_zh-hans.md
@@ -287,6 +287,7 @@ conda install -c huggingface transformers
 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。
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (来自 SHI Labs) 伴随论文 [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) 由 Ali Hassani and Humphrey Shi 发布。
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (来自 Meta AI) 伴随论文 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) 由 Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski 发布。
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (来自 HuggingFace), 伴随论文 [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) 由 Victor Sanh, Lysandre Debut and Thomas Wolf 发布。 同样的方法也应用于压缩 GPT-2 到 [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa 到 [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT 到 [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) 和德语版 DistilBERT。
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (来自 Microsoft Research) 伴随论文 [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) 由 Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei 发布。
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (来自 NAVER) 伴随论文 [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) 由 Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park 发布。

--- a/README_zh-hant.md
+++ b/README_zh-hant.md
@@ -299,6 +299,7 @@ conda install -c huggingface transformers
 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.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
+1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER) released with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.

--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@@ -472,6 +472,8 @@
        title: DETR
      - local: model_doc/dinat
        title: DiNAT
+      - local: model_doc/dinov2
+        title: DINO V2
      - local: model_doc/dit
        title: DiT
      - local: model_doc/dpt

--- a/docs/source/en/index.md
+++ b/docs/source/en/index.md
@@ -103,6 +103,7 @@ The documentation is organized into five sections:
 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.
 1. **[DiNAT](model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
+1. **[DINOv2](model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
 1. **[DiT](model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
@@ -323,6 +324,7 @@ Flax), PyTorch, and/or TensorFlow.
 |             DETA              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             DETR              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             DiNAT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            DINOv2             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          DistilBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |           DonutSwin           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |              DPR              |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |

--- a/docs/source/en/model_doc/dinov2.md
+++ b/docs/source/en/model_doc/dinov2.md
+<!--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.
+-->
+
+# DINOv2
+
+## Overview
+
+The DINOv2 model was proposed in [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by
+Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
+DINOv2 is an upgrade of [DINO](https://arxiv.org/abs/2104.14294), a self-supervised method applied on [Vision Transformers](vit). This method enables all-purpose visual features, i.e., features that work across image distributions and tasks without finetuning.
+
+The abstract from the paper is the following:
+
+*The recent breakthroughs in natural language processing for model pretraining on large quantities of data have opened the way for similar foundation models in computer vision. These models could greatly simplify the use of images in any system by producing all-purpose visual features, i.e., features that work across image distributions and tasks without finetuning. This work shows that existing pretraining methods, especially self-supervised methods, can produce such features if trained on enough curated data from diverse sources. We revisit existing approaches and combine different techniques to scale our pretraining in terms of data and model size. Most of the technical contributions aim at accelerating and stabilizing the training at scale. In terms of data, we propose an automatic pipeline to build a dedicated, diverse, and curated image dataset instead of uncurated data, as typically done in the self-supervised literature. In terms of models, we train a ViT model (Dosovitskiy et al., 2020) with 1B parameters and distill it into a series of smaller models that surpass the best available all-purpose features, OpenCLIP (Ilharco et al., 2021) on most of the benchmarks at image and pixel levels.*
+
+Tips:
+
+- One can use [`AutoImageProcessor`] class to prepare images for the model.
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/facebookresearch/dinov2).
+
+
+## Dinov2Config
+
+[[autodoc]] Dinov2Config
+
+## Dinov2Model
+
+[[autodoc]] Dinov2Model
+    - forward
+
+## Dinov2ForImageClassification
+
+[[autodoc]] Dinov2ForImageClassification
+    - forward
--- a/docs/source/en/tasks/image_classification.md
+++ b/docs/source/en/tasks/image_classification.md
@@ -34,7 +34,7 @@ The task illustrated in this tutorial is supported by the following model archit

 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->

-[BEiT](../model_doc/beit), [BiT](../model_doc/bit), [ConvNeXT](../model_doc/convnext), [ConvNeXTV2](../model_doc/convnextv2), [CvT](../model_doc/cvt), [Data2VecVision](../model_doc/data2vec-vision), [DeiT](../model_doc/deit), [DiNAT](../model_doc/dinat), [EfficientFormer](../model_doc/efficientformer), [EfficientNet](../model_doc/efficientnet), [FocalNet](../model_doc/focalnet), [ImageGPT](../model_doc/imagegpt), [LeViT](../model_doc/levit), [MobileNetV1](../model_doc/mobilenet_v1), [MobileNetV2](../model_doc/mobilenet_v2), [MobileViT](../model_doc/mobilevit), [MobileViTV2](../model_doc/mobilevitv2), [NAT](../model_doc/nat), [Perceiver](../model_doc/perceiver), [PoolFormer](../model_doc/poolformer), [RegNet](../model_doc/regnet), [ResNet](../model_doc/resnet), [SegFormer](../model_doc/segformer), [SwiftFormer](../model_doc/swiftformer), [Swin Transformer](../model_doc/swin), [Swin Transformer V2](../model_doc/swinv2), [VAN](../model_doc/van), [ViT](../model_doc/vit), [ViT Hybrid](../model_doc/vit_hybrid), [ViTMSN](../model_doc/vit_msn)
+[BEiT](../model_doc/beit), [BiT](../model_doc/bit), [ConvNeXT](../model_doc/convnext), [ConvNeXTV2](../model_doc/convnextv2), [CvT](../model_doc/cvt), [Data2VecVision](../model_doc/data2vec-vision), [DeiT](../model_doc/deit), [DiNAT](../model_doc/dinat), [DINOv2](../model_doc/dinov2), [EfficientFormer](../model_doc/efficientformer), [EfficientNet](../model_doc/efficientnet), [FocalNet](../model_doc/focalnet), [ImageGPT](../model_doc/imagegpt), [LeViT](../model_doc/levit), [MobileNetV1](../model_doc/mobilenet_v1), [MobileNetV2](../model_doc/mobilenet_v2), [MobileViT](../model_doc/mobilevit), [MobileViTV2](../model_doc/mobilevitv2), [NAT](../model_doc/nat), [Perceiver](../model_doc/perceiver), [PoolFormer](../model_doc/poolformer), [RegNet](../model_doc/regnet), [ResNet](../model_doc/resnet), [SegFormer](../model_doc/segformer), [SwiftFormer](../model_doc/swiftformer), [Swin Transformer](../model_doc/swin), [Swin Transformer V2](../model_doc/swinv2), [VAN](../model_doc/van), [ViT](../model_doc/vit), [ViT Hybrid](../model_doc/vit_hybrid), [ViTMSN](../model_doc/vit_msn)
 <!--End of the generated tip-->

 </Tip>

--- a/src/transformers/__init__.py
+++ b/src/transformers/__init__.py
@@ -293,6 +293,7 @@ _import_structure = {
    "models.detr": ["DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "DetrConfig"],
    "models.dialogpt": [],
    "models.dinat": ["DINAT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DinatConfig"],
+    "models.dinov2": ["DINOV2_PRETRAINED_CONFIG_ARCHIVE_MAP", "Dinov2Config"],
    "models.distilbert": ["DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DistilBertConfig", "DistilBertTokenizer"],
    "models.dit": [],
    "models.donut": ["DONUT_SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP", "DonutProcessor", "DonutSwinConfig"],
@@ -1579,6 +1580,14 @@ else:
            "DinatPreTrainedModel",
        ]
    )
+    _import_structure["models.dinov2"].extend(
+        [
+            "DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "Dinov2ForImageClassification",
+            "Dinov2Model",
+            "Dinov2PreTrainedModel",
+        ]
+    )
    _import_structure["models.distilbert"].extend(
        [
            "DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -4244,6 +4253,7 @@ if TYPE_CHECKING:
    from .models.deta import DETA_PRETRAINED_CONFIG_ARCHIVE_MAP, DetaConfig
    from .models.detr import DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, DetrConfig
    from .models.dinat import DINAT_PRETRAINED_CONFIG_ARCHIVE_MAP, DinatConfig
+    from .models.dinov2 import DINOV2_PRETRAINED_CONFIG_ARCHIVE_MAP, Dinov2Config
    from .models.distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig, DistilBertTokenizer
    from .models.donut import DONUT_SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP, DonutProcessor, DonutSwinConfig
    from .models.dpr import (
@@ -5360,6 +5370,12 @@ if TYPE_CHECKING:
            DinatModel,
            DinatPreTrainedModel,
        )
+        from .models.dinov2 import (
+            DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Dinov2ForImageClassification,
+            Dinov2Model,
+            Dinov2PreTrainedModel,
+        )
        from .models.distilbert import (
            DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
            DistilBertForMaskedLM,

--- a/src/transformers/models/__init__.py
+++ b/src/transformers/models/__init__.py
@@ -65,6 +65,7 @@ from . import (
    detr,
    dialogpt,
    dinat,
+    dinov2,
    distilbert,
    dit,
    donut,

--- a/src/transformers/models/auto/configuration_auto.py
+++ b/src/transformers/models/auto/configuration_auto.py
@@ -75,6 +75,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
        ("deta", "DetaConfig"),
        ("detr", "DetrConfig"),
        ("dinat", "DinatConfig"),
+        ("dinov2", "Dinov2Config"),
        ("distilbert", "DistilBertConfig"),
        ("donut-swin", "DonutSwinConfig"),
        ("dpr", "DPRConfig"),
@@ -276,6 +277,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
        ("deta", "DETA_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("detr", "DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("dinat", "DINAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("dinov2", "DINOV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("distilbert", "DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("donut-swin", "DONUT_SWIN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("dpr", "DPR_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -470,6 +472,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
        ("detr", "DETR"),
        ("dialogpt", "DialoGPT"),
        ("dinat", "DiNAT"),
+        ("dinov2", "DINOv2"),
        ("distilbert", "DistilBERT"),
        ("dit", "DiT"),
        ("donut-swin", "DonutSwin"),

--- a/src/transformers/models/auto/image_processing_auto.py
+++ b/src/transformers/models/auto/image_processing_auto.py
@@ -56,6 +56,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
        ("deta", "DetaImageProcessor"),
        ("detr", "DetrImageProcessor"),
        ("dinat", "ViTImageProcessor"),
+        ("dinov2", "BitImageProcessor"),
        ("donut-swin", "DonutImageProcessor"),
        ("dpt", "DPTImageProcessor"),
        ("efficientformer", "EfficientFormerImageProcessor"),

--- a/src/transformers/models/auto/modeling_auto.py
+++ b/src/transformers/models/auto/modeling_auto.py
@@ -73,6 +73,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
        ("deta", "DetaModel"),
        ("detr", "DetrModel"),
        ("dinat", "DinatModel"),
+        ("dinov2", "Dinov2Model"),
        ("distilbert", "DistilBertModel"),
        ("donut-swin", "DonutSwinModel"),
        ("dpr", "DPRQuestionEncoder"),
@@ -455,6 +456,7 @@ MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
        ("data2vec-vision", "Data2VecVisionForImageClassification"),
        ("deit", ("DeiTForImageClassification", "DeiTForImageClassificationWithTeacher")),
        ("dinat", "DinatForImageClassification"),
+        ("dinov2", "Dinov2ForImageClassification"),
        (
            "efficientformer",
            (

--- a/src/transformers/models/bit/image_processing_bit.py
+++ b/src/transformers/models/bit/image_processing_bit.py
@@ -76,10 +76,10 @@ class BitImageProcessor(BaseImageProcessor):
            method.
        do_normalize:
            Whether to normalize the image. Can be overridden by `do_normalize` in the `preprocess` method.
-        image_mean (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_MEAN`):
+        image_mean (`float` or `List[float]`, *optional*, defaults to `OPENAI_CLIP_MEAN`):
            Mean to use if normalizing the image. This is a float or list of floats the length of the number of
            channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method.
-        image_std (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_STD`):
+        image_std (`float` or `List[float]`, *optional*, defaults to `OPENAI_CLIP_MEAN`):
            Image standard deviation.
        do_convert_rgb (`bool`, *optional*, defaults to `True`):
            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the

--- a/src/transformers/models/dinov2/__init__.py
+++ b/src/transformers/models/dinov2/__init__.py
+# 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_dinov2": ["DINOV2_PRETRAINED_CONFIG_ARCHIVE_MAP", "Dinov2Config", "Dinov2OnnxConfig"]
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_dinov2"] = [
+        "DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "Dinov2ForImageClassification",
+        "Dinov2Model",
+        "Dinov2PreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_dinov2 import DINOV2_PRETRAINED_CONFIG_ARCHIVE_MAP, Dinov2Config, Dinov2OnnxConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_dinov2 import (
+            DINOV2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Dinov2ForImageClassification,
+            Dinov2Model,
+            Dinov2PreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
--- a/src/transformers/models/dinov2/configuration_dinov2.py
+++ b/src/transformers/models/dinov2/configuration_dinov2.py
+# 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.
+""" DINOv2 model configuration"""
+
+from collections import OrderedDict
+from typing import Mapping
+
+from packaging import version
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DINOV2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/dinov2-base": "https://huggingface.co/facebook/dinov2-base/resolve/main/config.json",
+}
+
+
+class Dinov2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Dinov2Model`]. It is used to instantiate an
+    Dinov2 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 Dinov2
+    [google/dinov2-base-patch16-224](https://huggingface.co/google/dinov2-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 the hidden size of the MLPs relative to the `hidden_size`.
+        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.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        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.
+        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.
+        layerscale_value (`float`, *optional*, defaults to 1.0):
+           Initial value to use for layer scale.
+        drop_path_rate (`float`, *optional*, defaults to 0.0):
+            Stochastic depth rate per sample (when applied in the main path of residual layers).
+        use_swiglu_ffn (`bool`, *optional*, defaults to `False`):
+            Whether to use the SwiGLU feedforward neural network.
+
+    Example:
+
+    ```python
+    >>> from transformers import Dinov2Config, Dinov2Model
+
+    >>> # Initializing a Dinov2 dinov2-base-patch16-224 style configuration
+    >>> configuration = Dinov2Config()
+
+    >>> # Initializing a model (with random weights) from the dinov2-base-patch16-224 style configuration
+    >>> model = Dinov2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "dinov2"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        mlp_ratio=4,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-6,
+        image_size=224,
+        patch_size=16,
+        num_channels=3,
+        qkv_bias=True,
+        layerscale_value=1.0,
+        drop_path_rate=0.0,
+        use_swiglu_ffn=False,
+        **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.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.qkv_bias = qkv_bias
+        self.layerscale_value = layerscale_value
+        self.drop_path_rate = drop_path_rate
+        self.use_swiglu_ffn = use_swiglu_ffn
+
+
+class Dinov2OnnxConfig(OnnxConfig):
+    torch_onnx_minimum_version = version.parse("1.11")
+
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        return OrderedDict(
+            [
+                ("pixel_values", {0: "batch", 1: "num_channels", 2: "height", 3: "width"}),
+            ]
+        )
+
+    @property
+    def atol_for_validation(self) -> float:
+        return 1e-4
--- a/src/transformers/models/dinov2/convert_dinov2_to_hf.py
+++ b/src/transformers/models/dinov2/convert_dinov2_to_hf.py
+# coding=utf-8
+# Copyright 2023 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 DINOv2 checkpoints from the original repository.
+
+URL: https://github.com/facebookresearch/dinov2/tree/main
+"""
+
+
+import argparse
+from pathlib import Path
+
+import requests
+import torch
+from PIL import Image
+from torchvision import transforms
+
+from transformers import BitImageProcessor, Dinov2Config, Dinov2Model
+from transformers.image_utils import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD, PILImageResampling
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_dinov2_config(model_name):
+    config = Dinov2Config(image_size=518, patch_size=14)
+
+    # size of the architecture
+    if "vits" in model_name:
+        config.hidden_size = 384
+        config.num_attention_heads = 6
+    elif "vitb" in model_name:
+        pass
+    elif "vitl" in model_name:
+        config.hidden_size = 1024
+        config.num_hidden_layers = 24
+        config.num_attention_heads = 16
+    elif "vitg" in model_name:
+        config.use_swiglu_ffn = True
+        config.hidden_size = 1536
+        config.num_hidden_layers = 40
+        config.num_attention_heads = 24
+    else:
+        raise ValueError("Model not supported")
+
+    return config
+
+
+def create_rename_keys(config):
+    rename_keys = []
+    # fmt: off
+
+    # patch embedding layer
+    rename_keys.append(("cls_token", "embeddings.cls_token"))
+    rename_keys.append(("mask_token", "embeddings.mask_token"))
+    rename_keys.append(("pos_embed", "embeddings.position_embeddings"))
+    rename_keys.append(("patch_embed.proj.weight", "embeddings.patch_embeddings.projection.weight"))
+    rename_keys.append(("patch_embed.proj.bias", "embeddings.patch_embeddings.projection.bias"))
+
+    for i in range(config.num_hidden_layers):
+        # layernorms
+        rename_keys.append((f"blocks.{i}.norm1.weight", f"encoder.layer.{i}.norm1.weight"))
+        rename_keys.append((f"blocks.{i}.norm1.bias", f"encoder.layer.{i}.norm1.bias"))
+        rename_keys.append((f"blocks.{i}.norm2.weight", f"encoder.layer.{i}.norm2.weight"))
+        rename_keys.append((f"blocks.{i}.norm2.bias", f"encoder.layer.{i}.norm2.bias"))
+        # MLP
+        if config.use_swiglu_ffn:
+            rename_keys.append((f"blocks.{i}.mlp.w12.weight", f"encoder.layer.{i}.mlp.w12.weight"))
+            rename_keys.append((f"blocks.{i}.mlp.w12.bias", f"encoder.layer.{i}.mlp.w12.bias"))
+            rename_keys.append((f"blocks.{i}.mlp.w3.weight", f"encoder.layer.{i}.mlp.w3.weight"))
+            rename_keys.append((f"blocks.{i}.mlp.w3.bias", f"encoder.layer.{i}.mlp.w3.bias"))
+        else:
+            rename_keys.append((f"blocks.{i}.mlp.fc1.weight", f"encoder.layer.{i}.mlp.fc1.weight"))
+            rename_keys.append((f"blocks.{i}.mlp.fc1.bias", f"encoder.layer.{i}.mlp.fc1.bias"))
+            rename_keys.append((f"blocks.{i}.mlp.fc2.weight", f"encoder.layer.{i}.mlp.fc2.weight"))
+            rename_keys.append((f"blocks.{i}.mlp.fc2.bias", f"encoder.layer.{i}.mlp.fc2.bias"))
+        # layerscale
+        rename_keys.append((f"blocks.{i}.ls1.gamma", f"encoder.layer.{i}.layer_scale1.lambda1"))
+        rename_keys.append((f"blocks.{i}.ls2.gamma", f"encoder.layer.{i}.layer_scale2.lambda1"))
+        # attention projection layer
+        rename_keys.append((f"blocks.{i}.attn.proj.weight", f"encoder.layer.{i}.attention.output.dense.weight"))
+        rename_keys.append((f"blocks.{i}.attn.proj.bias", f"encoder.layer.{i}.attention.output.dense.bias"))
+
+    # final layernorm
+    rename_keys.append(("norm.weight", "layernorm.weight"))
+    rename_keys.append(("norm.bias", "layernorm.bias"))
+
+    # fmt: on
+    return rename_keys
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+# we split up the matrix of each encoder layer into queries, keys and values
+def read_in_q_k_v(state_dict, config):
+    for i in range(config.num_hidden_layers):
+        # read in weights + bias of input projection layer (in timm, this is a single matrix + bias)
+        in_proj_weight = state_dict.pop(f"blocks.{i}.attn.qkv.weight")
+        in_proj_bias = state_dict.pop(f"blocks.{i}.attn.qkv.bias")
+        # next, add query, keys and values (in that order) to the state dict
+        state_dict[f"encoder.layer.{i}.attention.attention.query.weight"] = in_proj_weight[: config.hidden_size, :]
+        state_dict[f"encoder.layer.{i}.attention.attention.query.bias"] = in_proj_bias[: config.hidden_size]
+        state_dict[f"encoder.layer.{i}.attention.attention.key.weight"] = in_proj_weight[
+            config.hidden_size : config.hidden_size * 2, :
+        ]
+        state_dict[f"encoder.layer.{i}.attention.attention.key.bias"] = in_proj_bias[
+            config.hidden_size : config.hidden_size * 2
+        ]
+        state_dict[f"encoder.layer.{i}.attention.attention.value.weight"] = in_proj_weight[-config.hidden_size :, :]
+        state_dict[f"encoder.layer.{i}.attention.attention.value.bias"] = in_proj_bias[-config.hidden_size :]
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    image = Image.open(requests.get(url, stream=True).raw)
+    return image
+
+
+@torch.no_grad()
+def convert_dinov2_checkpoint(model_name, pytorch_dump_folder_path, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to our DINOv2 structure.
+    """
+
+    # define default Dinov2 configuration
+    config = get_dinov2_config(model_name)
+
+    # load original model from torch hub
+    original_model = torch.hub.load("facebookresearch/dinov2", model_name)
+    original_model.eval()
+
+    # load state_dict of original model, remove and rename some keys
+    state_dict = original_model.state_dict()
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+    read_in_q_k_v(state_dict, config)
+
+    for key, val in state_dict.copy().items():
+        val = state_dict.pop(key)
+        if "w12" in key:
+            key = key.replace("w12", "weights_in")
+        if "w3" in key:
+            key = key.replace("w3", "weights_out")
+        state_dict[key] = val
+
+    # load HuggingFace model
+    model = Dinov2Model(config, add_pooling_layer=False).eval()
+    model.load_state_dict(state_dict)
+
+    # load image
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+    # preprocess image
+    transformations = transforms.Compose(
+        [
+            transforms.Resize(256, interpolation=transforms.InterpolationMode.BICUBIC),
+            transforms.CenterCrop(224),
+            transforms.ToTensor(),
+            transforms.Normalize(
+                mean=IMAGENET_DEFAULT_MEAN,  # these are RGB mean+std values
+                std=IMAGENET_DEFAULT_STD,  # across a large photo dataset.
+            ),
+        ]
+    )
+
+    original_pixel_values = transformations(image).unsqueeze(0)  # insert batch dimension
+
+    processor = BitImageProcessor(
+        size={"shortest_edge": 256},
+        resample=PILImageResampling.BICUBIC,
+        image_mean=IMAGENET_DEFAULT_MEAN,
+        image_std=IMAGENET_DEFAULT_STD,
+    )
+    pixel_values = processor(image, return_tensors="pt").pixel_values
+
+    assert torch.allclose(original_pixel_values, pixel_values)
+
+    with torch.no_grad():
+        outputs = model(pixel_values)
+        original_outputs = original_model(pixel_values)
+
+    # assert values
+    assert outputs.last_hidden_state[:, 0].shape == original_outputs.shape
+    assert torch.allclose(outputs.last_hidden_state[:, 0], original_outputs, atol=1e-3)
+    print("Looks ok!")
+
+    if pytorch_dump_folder_path is not None:
+        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+        print(f"Saving model {model_name} to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        print(f"Saving image processor to {pytorch_dump_folder_path}")
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        model_name_to_hf_name = {
+            "dinov2_vits14": "dinov2-small",
+            "dinov2_vitb14": "dinov2-base",
+            "dinov2_vitl14": "dinov2-large",
+            "dinov2_vitg14": "dinov2-giant",
+        }
+
+        name = model_name_to_hf_name[model_name]
+        model.push_to_hub(f"facebook/{name}")
+        processor.push_to_hub(f"facebook/{name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--model_name",
+        default="dinov2_vitb14",
+        type=str,
+        choices=["dinov2_vits14", "dinov2_vitb14", "dinov2_vitl14", "dinov2_vitg14"],
+        help="Name of the model you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
+    )
+    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_dinov2_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)