Add BiT + ViT hybrid (#20550)

* First draft

* More improvements

* Add backbone, first draft of ViT hybrid

* Add AutoBackbone

* More improvements

* Fix bug

* More improvements

* More improvements

* Convert ViT-hybrid

* More improvements

* add patch bit

* Fix style

* Improve code

* cleaned v1

* more cleaning

* more refactoring

* Improve models, add tests

* Add docs and tests

* Make more tests pass

* Improve default backbone config

* Update model_type

* Fix more tests

* Add more copied from statements

* More improvements

* Add push to hub to conversion scripts

* clean

* more cleanup

* clean

* replace to

* fix

* Update src/transformers/models/bit/configuration_bit.py

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fix base model prefix

* more cleaning

* get rid of stem

* clean

* replace flag

* Update src/transformers/models/bit/configuration_bit.py

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/bit/configuration_bit.py

Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* add check

* another check

* fix for hybrid vit

* final fix

* update config

* fix class name

* fix `make fix-copies`

* remove `use_activation`

* Update src/transformers/models/bit/configuration_bit.py

* rm unneeded file

* Add BiT image processor

* rm unneeded file

* add doc

* Add image processor to conversion script

* Add ViTHybrid image processor

* Add resources

* Move bit to correct position

* Fix auto mapping

* Rename hybrid to Hybrid

* Fix name in toctree

* Fix READMEs'

* Improve config

* Simplify GroupNormActivation layer

* fix test + make style

* Improve config

* Apply suggestions from code review

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* remove comment

* remove comment

* replace

* replace

* remove all conv_layer

* refactor norm_layer

* revert x

* add copied from

* last changes + integration tests

* make fixup

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix name

* fix message

* remove assert and refactor

* refactor + make fixup

* refactor - add  + sfety checker

* fix docstring + checkpoint names

* fix merge issues

* fix function name

* fix copies

* Apply suggestions from code review

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix model checkpoint

* fix doctest output

* vit name on doc

* fix name on doc

* fix small nits

* fixed integration tests

* final changes - slow tests pass

Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
Co-authored-by: younesbelkada <younesbelkada@gmail.com>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

README.md

@@ -274,6 +274,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -405,6 +406,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.

README_es.md

@@ -274,6 +274,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -405,6 +406,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.

README_hd.md

@@ -247,6 +247,7 @@ conda install -c huggingface transformers
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (गूगल रिसर्च से) साथ वाला पेपर [बिग बर्ड: ट्रांसफॉर्मर्स फॉर लॉन्गर सीक्वेंस](https://arxiv .org/abs/2007.14062) मंज़िल ज़हीर, गुरु गुरुगणेश, अविनावा दुबे, जोशुआ आइंस्ली, क्रिस अल्बर्टी, सैंटियागो ओंटानोन, फिलिप फाम, अनिरुद्ध रावुला, किफ़ान वांग, ली यांग, अमर अहमद द्वारा।
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (गूगल रिसर्च से) साथ में पेपर [बिग बर्ड: ट्रांसफॉर्मर्स फॉर लॉन्गर सीक्वेंस](https://arxiv.org/abs/2007.14062) मंज़िल ज़हीर, गुरु गुरुगणेश, अविनावा दुबे, जोशुआ आइंस्ली, क्रिस अल्बर्टी, सैंटियागो ओंटानन, फिलिप फाम द्वारा , अनिरुद्ध रावुला, किफ़ान वांग, ली यांग, अमर अहमद द्वारा पोस्ट किया गया।
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (फेसबुक से) साथ में कागज [एक ओपन-डोमेन चैटबॉट बनाने की विधि](https://arxiv.org /abs/2004.13637) स्टीफन रोलर, एमिली दीनन, नमन गोयल, दा जू, मैरी विलियमसन, यिनहान लियू, जिंग जू, मायल ओट, कर्ट शस्टर, एरिक एम। स्मिथ, वाई-लैन बॉरो, जेसन वेस्टन द्वारा।
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (फेसबुक से) साथ में पेपर [एक ओपन-डोमेन चैटबॉट बनाने की रेसिपी](https://arxiv .org/abs/2004.13637) स्टीफन रोलर, एमिली दीनन, नमन गोयल, दा जू, मैरी विलियमसन, यिनहान लियू, जिंग जू, मायल ओट, कर्ट शस्टर, एरिक एम स्मिथ, वाई-लैन बॉरो, जेसन वेस्टन द्वारा।
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/).
@@ -378,6 +379,7 @@ conda install -c huggingface transformers
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (NAVER AI Lab/Kakao Enterprise/Kakao Brain से) साथ में कागज [ViLT: Vision-and-Language Transformer बिना कनवल्शन या रीजन सुपरविजन](https://arxiv.org/abs/2102.03334) वोनजे किम, बोक्यूंग सोन, इल्डू किम द्वारा पोस्ट किया गया।
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (फेसबुक एआई से) साथ में पेपर [wav2vec 2.0: ए फ्रेमवर्क फॉर सेल्फ-सुपरवाइज्ड लर्निंग ऑफ स्पीच रिप्रेजेंटेशन] (https://arxiv.org/abs/2006.11477) एलेक्सी बेवस्की, हेनरी झोउ, अब्देलरहमान मोहम्मद, माइकल औली द्वारा।

README_ja.md

@@ -309,6 +309,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -440,6 +441,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.

README_ko.md

@@ -224,6 +224,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -355,6 +356,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.

README_zh-hans.md

@@ -248,6 +248,7 @@ conda install -c huggingface transformers
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (来自 Google Research) 伴随论文 [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) 由 Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed 发布。
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (来自 Google Research) 伴随论文 [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) 由 Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed 发布。
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (来自 Microsoft Research AI4Science) 伴随论文 [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) 由 Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu 发布。
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (来自 Google AI) 伴随论文 [Big Transfer (BiT) 由 Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby 发布。
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston 发布。
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston 发布。
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -379,6 +380,7 @@ conda install -c huggingface transformers
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (来自 NAVER AI Lab/Kakao Enterprise/Kakao Brain) 伴随论文 [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) 由 Wonjae Kim, Bokyung Son, Ildoo Kim 发布。
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (来自 Facebook AI) 伴随论文 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 由 Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 发布。

README_zh-hant.md

@@ -260,6 +260,7 @@ conda install -c huggingface transformers
 1. **[BigBird-Pegasus](https://huggingface.co/docs/transformers/model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BioGpt](https://huggingface.co/docs/transformers/main/model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](https://huggingface.co/docs/transformers/main/model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -391,6 +392,7 @@ conda install -c huggingface transformers
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
 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/main/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. **[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. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.

docs/source/en/_toctree.yml

@@ -388,6 +388,8 @@
       sections:
       - local: model_doc/beit
         title: BEiT
+      - local: model_doc/bit
+        title: BiT
       - local: model_doc/conditional_detr
         title: Conditional DETR
       - local: model_doc/convnext
@@ -444,6 +446,8 @@
         title: VideoMAE
       - local: model_doc/vit
         title: Vision Transformer (ViT)
+      - local: model_doc/vit_hybrid
+        title: ViT Hybrid
       - local: model_doc/vit_mae
         title: ViTMAE
       - local: model_doc/vit_msn

docs/source/en/index.mdx

@@ -61,6 +61,7 @@ The documentation is organized into five sections:
 1. **[BigBird-Pegasus](model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BigBird-RoBERTa](model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[BioGpt](model_doc/biogpt)** (from Microsoft Research AI4Science) released with the paper [BioGPT: generative pre-trained transformer for biomedical text generation and mining](https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac409/6713511?guestAccessKey=a66d9b5d-4f83-4017-bb52-405815c907b9) by Renqian Luo, Liai Sun, Yingce Xia, Tao Qin, Sheng Zhang, Hoifung Poon and Tie-Yan Liu.
+1. **[BiT](model_doc/bit)** (from Google AI) released with the paper [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
 1. **[Blenderbot](model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
@@ -192,6 +193,7 @@ The documentation is organized into five sections:
 1. **[ViLT](model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
 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. **[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. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
@@ -231,6 +233,7 @@ Flax), PyTorch, and/or TensorFlow.
 |            BigBird            |       ✅       |       ✅       |       ✅        |         ❌         |      ✅      |
 |        BigBird-Pegasus        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            BioGpt             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|              BiT              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          Blenderbot           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |        BlenderbotSmall        |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             BLOOM             |       ❌       |       ✅       |       ✅        |         ❌         |      ❌      |
@@ -355,6 +358,7 @@ Flax), PyTorch, and/or TensorFlow.
 |     VisionTextDualEncoder     |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 |          VisualBERT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |              ViT              |       ❌       |       ❌       |       ✅        |         ✅         |      ✅      |
+|          ViT Hybrid           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            ViTMAE             |       ❌       |       ❌       |       ✅        |         ✅         |      ❌      |
 |            ViTMSN             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           Wav2Vec2            |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/bit.mdx

@@ -0,0 +1,65 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Big Transfer (BiT)
+
+## Overview
+
+The BiT model was proposed in [Big Transfer (BiT): General Visual Representation Learning](https://arxiv.org/abs/1912.11370) by Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, Neil Houlsby.
+BiT is a simple recipe for scaling up pre-training of [ResNet](resnet)-like architectures (specifically, ResNetv2). The method results in significant improvements for transfer learning.
+
+The abstract from the paper is the following:
+
+*Transfer of pre-trained representations improves sample efficiency and simplifies hyperparameter tuning when training deep neural networks for vision. We revisit the paradigm of pre-training on large supervised datasets and fine-tuning the model on a target task. We scale up pre-training, and propose a simple recipe that we call Big Transfer (BiT). By combining a few carefully selected components, and transferring using a simple heuristic, we achieve strong performance on over 20 datasets. BiT performs well across a surprisingly wide range of data regimes -- from 1 example per class to 1M total examples. BiT achieves 87.5% top-1 accuracy on ILSVRC-2012, 99.4% on CIFAR-10, and 76.3% on the 19 task Visual Task Adaptation Benchmark (VTAB). On small datasets, BiT attains 76.8% on ILSVRC-2012 with 10 examples per class, and 97.0% on CIFAR-10 with 10 examples per class. We conduct detailed analysis of the main components that lead to high transfer performance.*
+
+Tips:
+
+- BiT models are equivalent to ResNetv2 in terms of architecture, except that: 1) all batch normalization layers are replaced by [group normalization](https://arxiv.org/abs/1803.08494),
+2) [weight standardization](https://arxiv.org/abs/1903.10520) is used for convolutional layers. The authors show that the combination of both is useful for training with large batch sizes, and has a significant
+impact on transfer learning.
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/google-research/big_transfer).
+
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with BiT.
+
+<PipelineTag pipeline="image-classification"/>
+
+- [`BitForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
+
+If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+## BitConfig
+
+[[autodoc]] BitConfig
+
+
+## BitImageProcessor
+
+[[autodoc]] BitImageProcessor
+    - preprocess
+
+
+## BitModel
+
+[[autodoc]] BitModel
+    - forward
+
+
+## BitForImageClassification
+
+[[autodoc]] BitForImageClassification
+    - forward
+

docs/source/en/model_doc/vit_hybrid.mdx

@@ -0,0 +1,68 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Hybrid Vision Transformer (ViT Hybrid)
+
+## Overview
+
+The hybrid Vision Transformer (ViT) model was proposed in [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. It's the first paper that successfully trains a Transformer encoder on ImageNet, attaining
+very good results compared to familiar convolutional architectures. ViT hybrid is a slight variant of the [plain Vision Transformer](vit),
+by leveraging a convolutional backbone (specifically, [BiT](bit)) whose features are used as initial "tokens" for the Transformer.
+
+
+The abstract from the paper is the following:
+
+*While the Transformer architecture has become the de-facto standard for natural language processing tasks, its
+applications to computer vision remain limited. In vision, attention is either applied in conjunction with
+convolutional networks, or used to replace certain components of convolutional networks while keeping their overall
+structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to
+sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of
+data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.),
+Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring
+substantially fewer computational resources to train.*
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code (written in JAX) can be
+found [here](https://github.com/google-research/vision_transformer).
+
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ViT Hybrid.
+
+<PipelineTag pipeline="image-classification"/>
+
+- [`ViTHybridForImageClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/image-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_classification.ipynb).
+
+If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+
+## ViTHybridConfig
+
+[[autodoc]] ViTHybridConfig
+
+## ViTHybridImageProcessor
+
+[[autodoc]] ViTHybridImageProcessor
+    - preprocess
+
+## ViTHybridModel
+
+[[autodoc]] ViTHybridModel
+    - forward
+
+## ViTHybridForImageClassification
+
+[[autodoc]] ViTHybridForImageClassification
+    - forward

src/transformers/__init__.py

@@ -161,6 +161,7 @@ _import_structure = {
         "BigBirdPegasusConfig",
     ],
     "models.biogpt": ["BIOGPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BioGptConfig", "BioGptTokenizer"],
+    "models.bit": ["BIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BitConfig"],
     "models.blenderbot": ["BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BlenderbotConfig", "BlenderbotTokenizer"],
     "models.blenderbot_small": [
         "BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP",
@@ -417,6 +418,7 @@ _import_structure = {
     "models.vision_text_dual_encoder": ["VisionTextDualEncoderConfig", "VisionTextDualEncoderProcessor"],
     "models.visual_bert": ["VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "VisualBertConfig"],
     "models.vit": ["VIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTConfig"],
+    "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.wav2vec2": [
@@ -745,6 +747,7 @@ else:
     _import_structure["image_transforms"] = ["rescale", "resize", "to_pil_image"]
     _import_structure["image_utils"] = ["ImageFeatureExtractionMixin"]
     _import_structure["models.beit"].extend(["BeitFeatureExtractor", "BeitImageProcessor"])
+    _import_structure["models.bit"].extend(["BitImageProcessor"])
     _import_structure["models.chinese_clip"].extend(["ChineseCLIPFeatureExtractor", "ChineseCLIPImageProcessor"])
     _import_structure["models.clip"].extend(["CLIPFeatureExtractor", "CLIPImageProcessor"])
     _import_structure["models.conditional_detr"].extend(
@@ -775,6 +778,7 @@ else:
     _import_structure["models.videomae"].extend(["VideoMAEFeatureExtractor", "VideoMAEImageProcessor"])
     _import_structure["models.vilt"].extend(["ViltFeatureExtractor", "ViltImageProcessor", "ViltProcessor"])
     _import_structure["models.vit"].extend(["ViTFeatureExtractor", "ViTImageProcessor"])
+    _import_structure["models.vit_hybrid"].extend(["ViTHybridImageProcessor"])
     _import_structure["models.yolos"].extend(["YolosFeatureExtractor", "YolosImageProcessor"])
 
 # Timm-backed objects
@@ -1055,6 +1059,15 @@ else:
             "BioGptPreTrainedModel",
         ]
     )
+    _import_structure["models.bit"].extend(
+        [
+            "BIT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "BitBackbone",
+            "BitForImageClassification",
+            "BitModel",
+            "BitPreTrainedModel",
+        ]
+    )
     _import_structure["models.blenderbot"].extend(
         [
             "BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2219,6 +2232,14 @@ else:
             "ViTPreTrainedModel",
         ]
     )
+    _import_structure["models.vit_hybrid"].extend(
+        [
+            "VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "ViTHybridForImageClassification",
+            "ViTHybridModel",
+            "ViTHybridPreTrainedModel",
+        ]
+    )
     _import_structure["models.vit_mae"].extend(
         [
             "VIT_MAE_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3402,6 +3423,7 @@ if TYPE_CHECKING:
     from .models.big_bird import BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP, BigBirdConfig
     from .models.bigbird_pegasus import BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, BigBirdPegasusConfig
     from .models.biogpt import BIOGPT_PRETRAINED_CONFIG_ARCHIVE_MAP, BioGptConfig, BioGptTokenizer
+    from .models.bit import BIT_PRETRAINED_CONFIG_ARCHIVE_MAP, BitConfig
     from .models.blenderbot import BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP, BlenderbotConfig, BlenderbotTokenizer
     from .models.blenderbot_small import (
         BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -3634,6 +3656,7 @@ if TYPE_CHECKING:
     from .models.vision_text_dual_encoder import VisionTextDualEncoderConfig, VisionTextDualEncoderProcessor
     from .models.visual_bert import VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, VisualBertConfig
     from .models.vit import VIT_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTConfig
+    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.wav2vec2 import (
@@ -3918,6 +3941,7 @@ if TYPE_CHECKING:
         from .image_transforms import rescale, resize, to_pil_image
         from .image_utils import ImageFeatureExtractionMixin
         from .models.beit import BeitFeatureExtractor, BeitImageProcessor
+        from .models.bit import BitImageProcessor
         from .models.chinese_clip import ChineseCLIPFeatureExtractor, ChineseCLIPImageProcessor
         from .models.clip import CLIPFeatureExtractor, CLIPImageProcessor
         from .models.conditional_detr import ConditionalDetrFeatureExtractor, ConditionalDetrImageProcessor
@@ -3944,6 +3968,7 @@ if TYPE_CHECKING:
         from .models.videomae import VideoMAEFeatureExtractor, VideoMAEImageProcessor
         from .models.vilt import ViltFeatureExtractor, ViltImageProcessor, ViltProcessor
         from .models.vit import ViTFeatureExtractor, ViTImageProcessor
+        from .models.vit_hybrid import ViTHybridImageProcessor
         from .models.yolos import YolosFeatureExtractor, YolosImageProcessor
 
     # Modeling
@@ -4182,6 +4207,13 @@ if TYPE_CHECKING:
             BioGptModel,
             BioGptPreTrainedModel,
         )
+        from .models.bit import (
+            BIT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            BitBackbone,
+            BitForImageClassification,
+            BitModel,
+            BitPreTrainedModel,
+        )
         from .models.blenderbot import (
             BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST,
             BlenderbotForCausalLM,
@@ -5124,6 +5156,12 @@ if TYPE_CHECKING:
             ViTModel,
             ViTPreTrainedModel,
         )
+        from .models.vit_hybrid import (
+            VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ViTHybridForImageClassification,
+            ViTHybridModel,
+            ViTHybridPreTrainedModel,
+        )
         from .models.vit_mae import (
             VIT_MAE_PRETRAINED_MODEL_ARCHIVE_LIST,
             ViTMAEForPreTraining,

src/transformers/models/__init__.py

@@ -31,6 +31,7 @@ from . import (
     big_bird,
     bigbird_pegasus,
     biogpt,
+    bit,
     blenderbot,
     blenderbot_small,
     bloom,
@@ -166,6 +167,7 @@ from . import (
     vision_text_dual_encoder,
     visual_bert,
     vit,
+    vit_hybrid,
     vit_mae,
     vit_msn,
     wav2vec2,

src/transformers/models/auto/configuration_auto.py

@@ -38,6 +38,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("big_bird", "BigBirdConfig"),
         ("bigbird_pegasus", "BigBirdPegasusConfig"),
         ("biogpt", "BioGptConfig"),
+        ("bit", "BitConfig"),
         ("blenderbot", "BlenderbotConfig"),
         ("blenderbot-small", "BlenderbotSmallConfig"),
         ("bloom", "BloomConfig"),
@@ -162,6 +163,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("vision-text-dual-encoder", "VisionTextDualEncoderConfig"),
         ("visual_bert", "VisualBertConfig"),
         ("vit", "ViTConfig"),
+        ("vit_hybrid", "ViTHybridConfig"),
         ("vit_mae", "ViTMAEConfig"),
         ("vit_msn", "ViTMSNConfig"),
         ("wav2vec2", "Wav2Vec2Config"),
@@ -191,6 +193,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("big_bird", "BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bigbird_pegasus", "BIGBIRD_PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("biogpt", "BIOGPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bit", "BIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("blenderbot", "BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("blenderbot-small", "BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bloom", "BLOOM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -302,6 +305,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("vilt", "VILT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("visual_bert", "VISUAL_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vit", "VIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("vit_hybrid", "VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vit_mae", "VIT_MAE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vit_msn", "VIT_MSN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("wav2vec2", "WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -334,6 +338,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("big_bird", "BigBird"),
         ("bigbird_pegasus", "BigBird-Pegasus"),
         ("biogpt", "BioGpt"),
+        ("bit", "BiT"),
         ("blenderbot", "Blenderbot"),
         ("blenderbot-small", "BlenderbotSmall"),
         ("bloom", "BLOOM"),
@@ -474,6 +479,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("vision-text-dual-encoder", "VisionTextDualEncoder"),
         ("visual_bert", "VisualBERT"),
         ("vit", "ViT"),
+        ("vit_hybrid", "ViT Hybrid"),
         ("vit_mae", "ViTMAE"),
         ("vit_msn", "ViTMSN"),
         ("wav2vec2", "Wav2Vec2"),

src/transformers/models/auto/image_processing_auto.py

@@ -38,6 +38,7 @@ logger = logging.get_logger(__name__)
 IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
     [
         ("beit", "BeitImageProcessor"),
+        ("bit", "BitImageProcessor"),
         ("chinese_clip", "ChineseCLIPImageProcessor"),
         ("clip", "CLIPImageProcessor"),
         ("clipseg", "ViTImageProcessor"),
@@ -79,6 +80,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("videomae", "VideoMAEImageProcessor"),
         ("vilt", "ViltImageProcessor"),
         ("vit", "ViTImageProcessor"),
+        ("vit_hybrid", "ViTHybridImageProcessor"),
         ("vit_mae", "ViTImageProcessor"),
         ("vit_msn", "ViTImageProcessor"),
         ("xclip", "CLIPImageProcessor"),

src/transformers/models/auto/modeling_auto.py

@@ -37,6 +37,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("big_bird", "BigBirdModel"),
         ("bigbird_pegasus", "BigBirdPegasusModel"),
         ("biogpt", "BioGptModel"),
+        ("bit", "BitModel"),
         ("blenderbot", "BlenderbotModel"),
         ("blenderbot-small", "BlenderbotSmallModel"),
         ("bloom", "BloomModel"),
@@ -156,6 +157,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("vision-text-dual-encoder", "VisionTextDualEncoderModel"),
         ("visual_bert", "VisualBertModel"),
         ("vit", "ViTModel"),
+        ("vit_hybrid", "ViTHybridModel"),
         ("vit_mae", "ViTMAEModel"),
         ("vit_msn", "ViTMSNModel"),
         ("wav2vec2", "Wav2Vec2Model"),
@@ -372,6 +374,7 @@ MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Image Classification mapping
         ("beit", "BeitForImageClassification"),
+        ("bit", "BitForImageClassification"),
         ("convnext", "ConvNextForImageClassification"),
         ("cvt", "CvtForImageClassification"),
         ("data2vec-vision", "Data2VecVisionForImageClassification"),
@@ -399,6 +402,7 @@ MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("swinv2", "Swinv2ForImageClassification"),
         ("van", "VanForImageClassification"),
         ("vit", "ViTForImageClassification"),
+        ("vit_hybrid", "ViTHybridForImageClassification"),
         ("vit_msn", "ViTMSNForImageClassification"),
     ]
 )
@@ -854,6 +858,7 @@ _MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_BACKBONE_MAPPING_NAMES = OrderedDict(
     [
         # Backbone mapping
+        ("bit", "BitBackbone"),
         ("maskformer-swin", "MaskFormerSwinBackbone"),
         ("resnet", "ResNetBackbone"),
     ]

src/transformers/models/bit/__init__.py

@@ -0,0 +1,78 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+# rely on isort to merge the imports
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+
+
+_import_structure = {"configuration_bit": ["BIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "BitConfig", "BitOnnxConfig"]}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_bit"] = [
+        "BIT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "BitForImageClassification",
+        "BitModel",
+        "BitPreTrainedModel",
+        "BitBackbone",
+    ]
+
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_bit"] = ["BitImageProcessor"]
+
+
+if TYPE_CHECKING:
+    from .configuration_bit import BIT_PRETRAINED_CONFIG_ARCHIVE_MAP, BitConfig, BitOnnxConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_bit import (
+            BIT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            BitBackbone,
+            BitForImageClassification,
+            BitModel,
+            BitPreTrainedModel,
+        )
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_bit import BitImageProcessor
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/bit/configuration_bit.py

@@ -0,0 +1,133 @@
+# coding=utf-8
+# Copyright 2022 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.
+""" BiT model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+BIT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "google/bit-50": "https://huggingface.co/google/bit-50/resolve/main/config.json",
+}
+
+
+class BitConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`BitModel`]. It is used to instantiate an BiT
+    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 BiT
+    [google/bit-50](https://huggingface.co/google/bit-50) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        embedding_size (`int`, *optional*, defaults to 64):
+            Dimensionality (hidden size) for the embedding layer.
+        hidden_sizes (`List[int]`, *optional*, defaults to `[256, 512, 1024, 2048]`):
+            Dimensionality (hidden size) at each stage.
+        depths (`List[int]`, *optional*, defaults to `[3, 4, 6, 3]`):
+            Depth (number of layers) for each stage.
+        layer_type (`str`, *optional*, defaults to `"preactivation"`):
+            The layer to use, it can be either `"preactivation"` or `"bottleneck"`.
+        hidden_act (`str`, *optional*, defaults to `"relu"`):
+            The non-linear activation function in each block. If string, `"gelu"`, `"relu"`, `"selu"` and `"gelu_new"`
+            are supported.
+        global_padding (`str`, *optional*):
+            Padding strategy to use for the convolutional layers. Can be either `"valid"`, `"same"`, or `None`.
+        num_groups (`int`, *optional*, defaults to `32`):
+            Number of groups used for the `BitGroupNormActivation` layers.
+        drop_path_rate (`float`, *optional*, defaults to 0.0):
+            The drop path rate for the stochastic depth.
+        embedding_dynamic_padding (`bool`, *optional*, defaults to `False`):
+            Whether or not to make use of dynamic padding for the embedding layer.
+        output_stride (`int`, *optional*, defaults to 32):
+            The output stride of the model.
+        width_factor (`int`, *optional*, defaults to 1):
+            The width factor for the model.
+        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).
+
+    Example:
+    ```python
+    >>> from transformers import BitConfig, BitModel
+
+    >>> # Initializing a BiT bit-50 style configuration
+    >>> configuration = BitConfig()
+
+    >>> # Initializing a model (with random weights) from the bit-50 style configuration
+    >>> model = BitModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+    model_type = "bit"
+    layer_types = ["preactivation", "bottleneck"]
+    supported_padding = ["SAME", "VALID"]
+
+    def __init__(
+        self,
+        num_channels=3,
+        embedding_size=64,
+        hidden_sizes=[256, 512, 1024, 2048],
+        depths=[3, 4, 6, 3],
+        layer_type="preactivation",
+        hidden_act="relu",
+        global_padding=None,
+        num_groups=32,
+        drop_path_rate=0.0,
+        embedding_dynamic_padding=False,
+        output_stride=32,
+        width_factor=1,
+        out_features=None,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        if layer_type not in self.layer_types:
+            raise ValueError(f"layer_type={layer_type} is not one of {','.join(self.layer_types)}")
+        if global_padding is not None:
+            if global_padding.upper() in self.supported_padding:
+                global_padding = global_padding.upper()
+            else:
+                raise ValueError(f"Padding strategy {global_padding} not supported")
+        self.num_channels = num_channels
+        self.embedding_size = embedding_size
+        self.hidden_sizes = hidden_sizes
+        self.depths = depths
+        self.layer_type = layer_type
+        self.hidden_act = hidden_act
+        self.global_padding = global_padding
+        self.num_groups = num_groups
+        self.drop_path_rate = drop_path_rate
+        self.embedding_dynamic_padding = embedding_dynamic_padding
+        self.output_stride = output_stride
+        self.width_factor = width_factor
+
+        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, len(depths) + 1)]
+        if out_features is not None:
+            if not isinstance(out_features, list):
+                raise ValueError("out_features should be a list")
+            for feature in out_features:
+                if feature not in self.stage_names:
+                    raise ValueError(
+                        f"Feature {feature} is not a valid feature name. Valid names are {self.stage_names}"
+                    )
+        self.out_features = out_features

src/transformers/models/bit/convert_bit_to_pytorch.py

@@ -0,0 +1,178 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert BiT checkpoints from the timm library."""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+
+import requests
+from huggingface_hub import hf_hub_download
+from timm import create_model
+from timm.data import resolve_data_config
+from timm.data.transforms_factory import create_transform
+from transformers import BitConfig, BitForImageClassification, BitImageProcessor
+from transformers.image_utils import PILImageResampling
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def get_config(model_name):
+    repo_id = "huggingface/label-files"
+    filename = "imagenet-1k-id2label.json"
+    id2label = json.load(open(hf_hub_download(repo_id, filename, repo_type="dataset"), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    label2id = {v: k for k, v in id2label.items()}
+
+    conv_layer = "std_conv" if "bit" in model_name else False
+
+    # note that when using BiT as backbone for ViT-hybrid checkpoints,
+    # one needs to additionally set config.layer_type = "bottleneck", config.stem_type = "same",
+    # config.conv_layer = "std_conv_same"
+    config = BitConfig(
+        conv_layer=conv_layer,
+        num_labels=1000,
+        id2label=id2label,
+        label2id=label2id,
+    )
+
+    return config
+
+
+def rename_key(name):
+    if "stem.conv" in name:
+        name = name.replace("stem.conv", "bit.embedder.convolution")
+    if "blocks" in name:
+        name = name.replace("blocks", "layers")
+    if "head.fc" in name:
+        name = name.replace("head.fc", "classifier.1")
+    if name.startswith("norm"):
+        name = "bit." + name
+    if "bit" not in name and "classifier" not in name:
+        name = "bit.encoder." + name
+
+    return name
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@torch.no_grad()
+def convert_bit_checkpoint(model_name, pytorch_dump_folder_path, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to our BiT structure.
+    """
+
+    # define default BiT configuration
+    config = get_config(model_name)
+
+    # load original model from timm
+    timm_model = create_model(model_name, pretrained=True)
+    timm_model.eval()
+
+    # load state_dict of original model
+    state_dict = timm_model.state_dict()
+    for key in state_dict.copy().keys():
+        val = state_dict.pop(key)
+        state_dict[rename_key(key)] = val.squeeze() if "head" in key else val
+
+    # load HuggingFace model
+    model = BitForImageClassification(config)
+    model.eval()
+    model.load_state_dict(state_dict)
+
+    # create image processor
+    transform = create_transform(**resolve_data_config({}, model=timm_model))
+    timm_transforms = transform.transforms
+
+    pillow_resamplings = {
+        "bilinear": PILImageResampling.BILINEAR,
+        "bicubic": PILImageResampling.BICUBIC,
+        "nearest": PILImageResampling.NEAREST,
+    }
+
+    processor = BitImageProcessor(
+        do_resize=True,
+        size={"shortest_edge": timm_transforms[0].size},
+        resample=pillow_resamplings[timm_transforms[0].interpolation.value],
+        do_center_crop=True,
+        crop_size={"height": timm_transforms[1].size[0], "width": timm_transforms[1].size[1]},
+        do_normalize=True,
+        image_mean=timm_transforms[-1].mean.tolist(),
+        image_std=timm_transforms[-1].std.tolist(),
+    )
+
+    image = prepare_img()
+    timm_pixel_values = transform(image).unsqueeze(0)
+    pixel_values = processor(image, return_tensors="pt").pixel_values
+
+    # verify pixel values
+    assert torch.allclose(timm_pixel_values, pixel_values)
+
+    # verify logits
+    with torch.no_grad():
+        outputs = model(pixel_values)
+        logits = outputs.logits
+
+    print("Logits:", logits[0, :3])
+    print("Predicted class:", model.config.id2label[logits.argmax(-1).item()])
+    timm_logits = timm_model(pixel_values)
+    assert timm_logits.shape == outputs.logits.shape
+    assert torch.allclose(timm_logits, outputs.logits, 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} and processor to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        print(f"Pushing model {model_name} and processor to the hub")
+        model.push_to_hub(f"ybelkada/{model_name}")
+        processor.push_to_hub(f"ybelkada/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--model_name",
+        default="resnetv2_50x1_bitm",
+        type=str,
+        help="Name of the BiT timm 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 to push the model to the hub.",
+    )
+
+    args = parser.parse_args()
+    convert_bit_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/bit/image_processing_bit.py

@@ -0,0 +1,345 @@
+# coding=utf-8
+# Copyright 2022 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.
+"""Image processor class for BiT."""
+
+from typing import Any, Dict, List, Optional, Union
+
+import numpy as np
+
+from transformers.utils.generic import TensorType
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from ...image_transforms import (
+    center_crop,
+    get_resize_output_image_size,
+    normalize,
+    rescale,
+    resize,
+    to_channel_dimension_format,
+)
+from ...image_utils import ChannelDimension, ImageInput, PILImageResampling, is_batched, to_numpy_array, valid_images
+from ...utils import logging
+from ...utils.import_utils import is_vision_available
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    import PIL
+
+
+def convert_to_rgb(image: Union[Any, PIL.Image.Image]) -> Union[Any, PIL.Image.Image]:
+    """
+    Converts `PIL.Image.Image` to RGB format. Images in other formats are returned as is.
+
+    Args:
+        image (`PIL.Image.Image`):
+            The image to convert.
+    """
+    if not isinstance(image, PIL.Image.Image):
+        return image
+
+    return image.convert("RGB")
+
+
+class BitImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a BiT image processor.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by
+            `do_resize` in the `preprocess` method.
+        size (`Dict[str, int]` *optional*, defaults to `{"shortest_edge": 224}`):
+            Size of the image after resizing. The shortest edge of the image is resized to size["shortest_edge"], with
+            the longest edge resized to keep the input aspect ratio. Can be overridden by `size` in the `preprocess`
+            method.
+        resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Can be overridden by `resample` in the `preprocess` method.
+        do_center_crop (`bool`, *optional*, defaults to `True`):
+            Whether to center crop the image to the specified `crop_size`. Can be overridden by `do_center_crop` in the
+            `preprocess` method.
+        crop_size (`Dict[str, int]` *optional*, defaults to 224):
+            Size of the output image after applying `center_crop`. Can be overridden by `crop_size` in the `preprocess`
+            method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
+            the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
+            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`):
+            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 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
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_center_crop: bool = True,
+        crop_size: Dict[str, int] = None,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        **kwargs
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"shortest_edge": 224}
+        size = get_size_dict(size, default_to_square=False)
+        crop_size = crop_size if crop_size is not None else {"height": 224, "width": 224}
+        crop_size = get_size_dict(crop_size, default_to_square=True, param_name="crop_size")
+
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
+        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.do_convert_rgb = do_convert_rgb
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Resize an image. The shortest edge of the image is resized to size["shortest_edge"], with the longest edge
+        resized to keep the input aspect ratio.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`Dict[str, int]`):
+                Size of the output image.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+                Resampling filter to use when resiizing the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        size = get_size_dict(size, default_to_square=False)
+        if "shortest_edge" not in size:
+            raise ValueError(f"The `size` parameter must contain the key `shortest_edge`. Got {size.keys()}")
+        output_size = get_resize_output_image_size(image, size=size["shortest_edge"], default_to_square=False)
+        return resize(image, size=output_size, resample=resample, data_format=data_format, **kwargs)
+
+    def center_crop(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Center crop an image. If the image is too small to be cropped to the size given, it will be padded (so the
+        returned result will always be of size `size`).
+
+        Args:
+            image (`np.ndarray`):
+                Image to center crop.
+            size (`Dict[str, int]`):
+                Size of the output image in the form of a dictionary with keys `height` and `width`.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        size = get_size_dict(size)
+        if "height" not in size or "width" not in size:
+            raise ValueError(f"The `size` parameter must contain the keys (height, width). Got {size.keys()}")
+        return center_crop(image, size=(size["height"], size["width"]), data_format=data_format, **kwargs)
+
+    def rescale(
+        self,
+        image: np.ndarray,
+        scale: Union[int, float],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ):
+        """
+        Rescale an image by a scale factor. image = image * scale.
+
+        Args:
+            image (`np.ndarray`):
+                Image to rescale.
+            scale (`int` or `float`):
+                Scale to apply to the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        return rescale(image, scale=scale, data_format=data_format, **kwargs)
+
+    def normalize(
+        self,
+        image: np.ndarray,
+        mean: Union[float, List[float]],
+        std: Union[float, List[float]],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Normalize an image. image = (image - image_mean) / image_std.
+
+        Args:
+            image (`np.ndarray`):
+                Image to normalize.
+            image_mean (`float` or `List[float]`):
+                Image mean.
+            image_std (`float` or `List[float]`):
+                Image standard deviation.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        return normalize(image, mean=mean, std=std, data_format=data_format, **kwargs)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: bool = None,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = None,
+        do_center_crop: bool = None,
+        crop_size: int = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        **kwargs
+    ) -> PIL.Image.Image:
+        """
+        Preprocess an image or batch of images.
+
+        Args:
+            images (`ImageInput`):
+                Image to preprocess.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. Shortest edge of the image is resized to size["shortest_edge"], with
+                the longest edge resized to keep the input aspect ratio.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_center_crop (`bool`, *optional*, defaults to `self.do_center_crop`):
+                Whether to center crop the image.
+            crop_size (`Dict[str, int]`, *optional*, defaults to `self.crop_size`):
+                Size of the center crop. Only has an effect if `do_center_crop` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: defaults to the channel dimension format of the input image.
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        size = size if size is not None else self.size
+        size = get_size_dict(size, param_name="size", default_to_square=False)
+        resample = resample if resample is not None else self.resample
+        do_center_crop = do_center_crop if do_center_crop is not None else self.do_center_crop
+        crop_size = crop_size if crop_size is not None else self.crop_size
+        crop_size = get_size_dict(crop_size, param_name="crop_size", default_to_square=True)
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        if not is_batched(images):
+            images = [images]
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        if do_resize and size is None:
+            raise ValueError("Size must be specified if do_resize is True.")
+
+        if do_center_crop and crop_size is None:
+            raise ValueError("Crop size must be specified if do_center_crop is True.")
+
+        if do_rescale and rescale_factor is None:
+            raise ValueError("Rescale factor must be specified if do_rescale is True.")
+
+        if do_normalize and (image_mean is None or image_std is None):
+            raise ValueError("Image mean and std must be specified if do_normalize is True.")
+
+        # PIL RGBA images are converted to RGB
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_resize:
+            images = [self.resize(image=image, size=size, resample=resample) for image in images]
+
+        if do_center_crop:
+            images = [self.center_crop(image=image, size=crop_size) for image in images]
+
+        if do_rescale:
+            images = [self.rescale(image=image, scale=rescale_factor) for image in images]
+
+        if do_normalize:
+            images = [self.normalize(image=image, mean=image_mean, std=image_std) for image in images]
+
+        images = [to_channel_dimension_format(image, data_format) for image in images]
+
+        data = {"pixel_values": images}
+        return BatchFeature(data=data, tensor_type=return_tensors)

src/transformers/models/bit/modeling_bit.py

@@ -0,0 +1,926 @@
+# coding=utf-8
+# Copyright 2022 Google 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 BiT model. Also supports backbone for ViT hybrid."""
+
+import collections
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import Tensor, nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BackboneOutput,
+    BaseModelOutputWithNoAttention,
+    BaseModelOutputWithPoolingAndNoAttention,
+    ImageClassifierOutputWithNoAttention,
+)
+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 .configuration_bit import BitConfig
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "BitConfig"
+_FEAT_EXTRACTOR_FOR_DOC = "AutoFeatureExtractor"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "google/bit-50"
+_EXPECTED_OUTPUT_SHAPE = [1, 2048, 7, 7]
+
+# Image classification docstring
+_IMAGE_CLASS_CHECKPOINT = "google/bit-50"
+_IMAGE_CLASS_EXPECTED_OUTPUT = "tiger cat"
+
+BIT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/bit-50",
+    # See all BiT models at https://huggingface.co/models?filter=bit
+]
+
+
+def get_padding_value(padding=None, kernel_size=7, stride=1, dilation=1) -> Tuple[Tuple, bool]:
+    r"""
+    Utility function to get the tuple padding value given the kernel_size and padding.
+
+    Args:
+        padding (Union[`str`, `int`], *optional*):
+            Padding value, can be either `"same"`, `"valid"`. If a different value is provided the default padding from
+            PyTorch is used.
+        kernel_size (`int`, *optional*, defaults to 7):
+            Kernel size of the convolution layers.
+        stride (`int`, *optional*, defaults to 1):
+            Stride value of the convolution layers.
+        dilation (`int`, *optional*, defaults to 1):
+            Dilation value of the convolution layers.
+    """
+    dynamic = False
+    if padding is None:
+        padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2
+        return padding, dynamic
+
+    if isinstance(padding, str):
+        # for any string padding, the padding will be calculated for you, one of three ways
+        padding = padding.lower()
+        if padding == "same":
+            # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact
+            if stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0:
+                # static case, no extra overhead
+                padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2
+            else:
+                # dynamic 'SAME' padding, has runtime/GPU memory overhead
+                padding = 0
+                dynamic = True
+        elif padding == "valid":
+            # 'VALID' padding, same as padding=0
+            padding = 0
+        else:
+            # Default to PyTorch style 'same'-ish symmetric padding
+            padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2
+    return padding, dynamic
+
+
+class WeightStandardizedConv2d(nn.Conv2d):
+    """Conv2d with Weight Standardization. Includes TensorFlow compatible SAME padding. Used for ViT Hybrid model.
+
+    Paper: [Micro-Batch Training with Batch-Channel Normalization and Weight
+    Standardization](https://arxiv.org/abs/1903.10520v2)
+    """
+
+    def __init__(
+        self,
+        in_channel,
+        out_channels,
+        kernel_size,
+        stride=1,
+        padding="SAME",
+        dilation=1,
+        groups=1,
+        bias=False,
+        eps=1e-6,
+    ):
+        padding, is_dynamic = get_padding_value(padding, kernel_size, stride=stride, dilation=dilation)
+        super().__init__(
+            in_channel,
+            out_channels,
+            kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=bias,
+        )
+        if is_dynamic:
+            self.pad = DynamicPad2d(kernel_size, stride, dilation)
+        else:
+            self.pad = None
+        self.eps = eps
+
+    def forward(self, hidden_state):
+        if self.pad is not None:
+            hidden_state = self.pad(hidden_state)
+        weight = nn.functional.batch_norm(
+            self.weight.reshape(1, self.out_channels, -1), None, None, training=True, momentum=0.0, eps=self.eps
+        ).reshape_as(self.weight)
+        hidden_state = nn.functional.conv2d(
+            hidden_state, weight, self.bias, self.stride, self.padding, self.dilation, self.groups
+        )
+        return hidden_state
+
+
+class BitGroupNormActivation(nn.GroupNorm):
+    r"""
+    A module that combines group normalization with an activation function.
+    """
+
+    def __init__(self, config, num_channels, eps=1e-5, affine=True, apply_activation=True):
+        super(BitGroupNormActivation, self).__init__(config.num_groups, num_channels, eps=eps, affine=affine)
+        if apply_activation:
+            self.activation = ACT2FN[config.hidden_act]
+        else:
+            self.activation = nn.Identity()
+
+    def forward(self, hidden_state):
+        hidden_state = nn.functional.group_norm(hidden_state, self.num_groups, self.weight, self.bias, self.eps)
+        hidden_state = self.activation(hidden_state)
+        return hidden_state
+
+
+class DynamicPad2d(nn.Module):
+    r"""
+    A module that wraps dynamic padding of any input, given the parameters of the convolutional layer and the input
+    hidden states.
+    """
+
+    def __init__(self, kernel_size, stride, dilation, value=0):
+        super().__init__()
+        # Safety checkers
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size, kernel_size)
+
+        if isinstance(stride, int):
+            stride = (stride, stride)
+
+        if isinstance(dilation, int):
+            dilation = (dilation, dilation)
+
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.dilation = dilation
+        self.value = value
+
+        def compute_padding(x, kernel_size, stride, dilation):
+            return max((math.ceil(x / stride) - 1) * stride + (kernel_size - 1) * dilation + 1 - x, 0)
+
+        self.compute_padding = compute_padding
+
+    def __call__(self, input):
+        # Get width and height
+        input_height, input_width = input.size()[-2:]
+
+        # Compute the padding values
+        padding_height = self.compute_padding(input_height, self.kernel_size[0], self.stride[0], self.dilation[0])
+        padding_width = self.compute_padding(input_width, self.kernel_size[1], self.stride[1], self.dilation[1])
+
+        # apply pad
+        if padding_height > 0 or padding_width > 0:
+            input = nn.functional.pad(
+                input,
+                [
+                    padding_width // 2,
+                    padding_width - padding_width // 2,
+                    padding_height // 2,
+                    padding_height - padding_height // 2,
+                ],
+                value=self.value,
+            )
+        return input
+
+
+class BitMaxPool2d(nn.MaxPool2d):
+    """Tensorflow like 'SAME' wrapper for 2D max pooling"""
+
+    def __init__(
+        self,
+        kernel_size: int,
+        stride=None,
+        dilation=1,
+        ceil_mode=False,
+        padding=(0, 0),
+        padding_value=0,
+        use_dynamic_padding=True,
+    ):
+        kernel_size = kernel_size if isinstance(kernel_size, collections.abc.Iterable) else (kernel_size, kernel_size)
+        stride = stride if isinstance(stride, collections.abc.Iterable) else (stride, stride)
+        dilation = dilation if isinstance(dilation, collections.abc.Iterable) else (dilation, dilation)
+        super().__init__(kernel_size, stride, padding, dilation, ceil_mode)
+        if use_dynamic_padding:
+            self.pad = DynamicPad2d(kernel_size, stride, dilation, padding_value)
+        else:
+            self.pad = nn.Identity()
+
+    def forward(self, hidden_states):
+        hidden_states = self.pad(hidden_states)
+        return nn.functional.max_pool2d(
+            hidden_states, self.kernel_size, self.stride, self.padding, self.dilation, self.ceil_mode
+        )
+
+
+class BitEmbeddings(nn.Module):
+    """
+    BiT Embeddings (stem) composed of a single aggressive convolution.
+    """
+
+    def __init__(self, config: BitConfig):
+        super().__init__()
+
+        self.convolution = WeightStandardizedConv2d(
+            config.num_channels,
+            config.embedding_size,
+            kernel_size=7,
+            stride=2,
+            eps=1e-8,
+            padding=config.global_padding,
+        )
+
+        self.pooler = BitMaxPool2d(kernel_size=3, stride=2, use_dynamic_padding=config.embedding_dynamic_padding)
+
+        # Use the same padding strategy as convolutional layers
+        if config.global_padding is not None and config.global_padding.upper() == "SAME":
+            self.pad = nn.Identity()
+        else:
+            self.pad = nn.ConstantPad2d(padding=(1, 1, 1, 1), value=0.0)
+
+        if not config.layer_type == "preactivation":
+            self.norm = BitGroupNormActivation(config, num_channels=config.embedding_size)
+        else:
+            self.norm = nn.Identity()
+
+        self.num_channels = config.num_channels
+
+    def forward(self, pixel_values: Tensor) -> Tensor:
+        num_channels = pixel_values.shape[1]
+        if num_channels != self.num_channels:
+            raise ValueError(
+                "Make sure that the channel dimension of the pixel values match with the one set in the configuration."
+            )
+
+        embedding = self.convolution(pixel_values)
+
+        embedding = self.pad(embedding)
+
+        embedding = self.norm(embedding)
+
+        embedding = self.pooler(embedding)
+
+        return embedding
+
+
+# Copied from transformers.models.convnext.modeling_convnext.drop_path
+def drop_path(input, drop_prob: float = 0.0, training: bool = False):
+    """
+    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 with Beit->Bit
+class BitDropPath(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)
+
+
+def make_div(value, divisor=8):
+    min_value = divisor
+    new_value = max(min_value, int(value + divisor / 2) // divisor * divisor)
+    if new_value < 0.9 * value:
+        new_value += divisor
+    return new_value
+
+
+class BitPreActivationBottleneckLayer(nn.Module):
+    """Pre-activation (v2) bottleneck block.
+    Follows the implementation of "Identity Mappings in Deep Residual Networks":
+    https://github.com/KaimingHe/resnet-1k-layers/blob/master/resnet-pre-act.lua
+
+    Except it puts the stride on 3x3 conv when available.
+    """
+
+    def __init__(
+        self,
+        config,
+        in_channels,
+        out_channels=None,
+        bottle_ratio=0.25,
+        stride=1,
+        dilation=1,
+        first_dilation=None,
+        groups=1,
+        drop_path_rate=0.0,
+        is_first_layer=False,
+    ):
+        super().__init__()
+
+        first_dilation = first_dilation or dilation
+
+        out_channels = out_channels or in_channels
+        mid_channels = make_div(out_channels * bottle_ratio)
+
+        if is_first_layer:
+            self.downsample = BitDownsampleConv(
+                config,
+                in_channels,
+                out_channels,
+                stride=stride,
+                preact=True,
+            )
+        else:
+            self.downsample = None
+
+        self.norm1 = BitGroupNormActivation(config, in_channels)
+        self.conv1 = WeightStandardizedConv2d(in_channels, mid_channels, 1, eps=1e-8, padding=config.global_padding)
+
+        self.norm2 = BitGroupNormActivation(config, num_channels=mid_channels)
+        self.conv2 = WeightStandardizedConv2d(
+            mid_channels, mid_channels, 3, stride=stride, groups=groups, eps=1e-8, padding=config.global_padding
+        )
+
+        self.norm3 = BitGroupNormActivation(config, mid_channels)
+        self.conv3 = WeightStandardizedConv2d(mid_channels, out_channels, 1, eps=1e-8, padding=config.global_padding)
+
+        self.drop_path = BitDropPath(drop_path_rate) if drop_path_rate > 0 else nn.Identity()
+
+    def forward(self, hidden_states):
+        hidden_states_preact = self.norm1(hidden_states)
+
+        # shortcut branch
+        shortcut = hidden_states
+        if self.downsample is not None:
+            shortcut = self.downsample(hidden_states_preact)
+
+        # residual branch
+        hidden_states = self.conv1(hidden_states_preact)
+        hidden_states = self.conv2(self.norm2(hidden_states))
+        hidden_states = self.conv3(self.norm3(hidden_states))
+        hidden_states = self.drop_path(hidden_states)
+        return hidden_states + shortcut
+
+
+class BitBottleneckLayer(nn.Module):
+    """Non Pre-activation bottleneck block, equivalent to V1.5/V1b bottleneck. Used for ViT Hybrid."""
+
+    def __init__(
+        self,
+        config,
+        in_channels,
+        out_channels=None,
+        bottle_ratio=0.25,
+        stride=1,
+        dilation=1,
+        first_dilation=None,
+        groups=1,
+        drop_path_rate=0.0,
+        is_first_layer=False,
+    ):
+        super().__init__()
+        first_dilation = first_dilation or dilation
+
+        out_channels = out_channels or in_channels
+        mid_chs = make_div(out_channels * bottle_ratio)
+
+        if is_first_layer:
+            self.downsample = BitDownsampleConv(
+                config,
+                in_channels,
+                out_channels,
+                stride=stride,
+                preact=False,
+            )
+        else:
+            self.downsample = None
+
+        self.conv1 = WeightStandardizedConv2d(in_channels, mid_chs, 1, eps=1e-8, padding=config.global_padding)
+        self.norm1 = BitGroupNormActivation(config, num_channels=mid_chs)
+        self.conv2 = WeightStandardizedConv2d(
+            mid_chs,
+            mid_chs,
+            3,
+            stride=stride,
+            dilation=first_dilation,
+            groups=groups,
+            eps=1e-8,
+            padding=config.global_padding,
+        )
+        self.norm2 = BitGroupNormActivation(config, num_channels=mid_chs)
+        self.conv3 = WeightStandardizedConv2d(mid_chs, out_channels, 1, eps=1e-8, padding=config.global_padding)
+        self.norm3 = BitGroupNormActivation(config, num_channels=out_channels, apply_activation=False)
+        self.drop_path = BitDropPath(drop_path_rate) if drop_path_rate > 0 else nn.Identity()
+
+        self.activation = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_states):
+        # shortcut branch
+        shortcut = hidden_states
+        if self.downsample is not None:
+            shortcut = self.downsample(hidden_states)
+
+        # residual
+        hidden_states = self.conv1(hidden_states)
+        hidden_states = self.norm1(hidden_states)
+
+        hidden_states = self.conv2(hidden_states)
+        hidden_states = self.norm2(hidden_states)
+
+        hidden_states = self.conv3(hidden_states)
+        hidden_states = self.norm3(hidden_states)
+
+        hidden_states = self.drop_path(hidden_states)
+        hidden_states = self.activation(hidden_states + shortcut)
+        return hidden_states
+
+
+class BitDownsampleConv(nn.Module):
+    def __init__(
+        self,
+        config,
+        in_channels,
+        out_channels,
+        stride=1,
+        preact=True,
+    ):
+        super().__init__()
+        self.conv = WeightStandardizedConv2d(
+            in_channels, out_channels, 1, stride=stride, eps=1e-8, padding=config.global_padding
+        )
+        self.norm = (
+            nn.Identity()
+            if preact
+            else BitGroupNormActivation(config, num_channels=out_channels, apply_activation=False)
+        )
+
+    def forward(self, x):
+        return self.norm(self.conv(x))
+
+
+class BitStage(nn.Module):
+    """
+    A ResNet v2 stage composed by stacked layers.
+    """
+
+    def __init__(
+        self,
+        config,
+        in_channels,
+        out_channels,
+        stride,
+        dilation,
+        depth,
+        bottle_ratio=0.25,
+        layer_dropout=None,
+    ):
+        super().__init__()
+
+        first_dilation = 1 if dilation in (1, 2) else 2
+
+        # Get the layer type
+        if config.layer_type == "bottleneck":
+            layer_cls = BitBottleneckLayer
+        else:
+            layer_cls = BitPreActivationBottleneckLayer
+
+        prev_chs = in_channels
+        self.layers = nn.Sequential()
+        for layer_idx in range(depth):
+            # Get the current hyper-parameters
+            stride, drop_path_rate, is_first_layer = self._get_updated_hyperparameters(
+                layer_idx, stride, layer_dropout
+            )
+
+            self.layers.add_module(
+                str(layer_idx),
+                layer_cls(
+                    config,
+                    prev_chs,
+                    out_channels,
+                    stride=stride,
+                    dilation=dilation,
+                    bottle_ratio=bottle_ratio,
+                    first_dilation=first_dilation,
+                    drop_path_rate=drop_path_rate,
+                    is_first_layer=is_first_layer,
+                ),
+            )
+            prev_chs = out_channels
+            first_dilation = dilation
+
+    def _get_updated_hyperparameters(self, layer_idx, stride, layer_dropout):
+        r"""
+        Get the new hyper-parameters with respect to the previous ones and the index of the current layer.
+        """
+        if layer_dropout:
+            drop_path_rate = layer_dropout[layer_idx]
+        else:
+            drop_path_rate = 0.0
+
+        if layer_idx != 0:
+            stride = 1
+
+        is_first_layer = layer_idx == 0
+
+        return stride, drop_path_rate, is_first_layer
+
+    def forward(self, input: Tensor) -> Tensor:
+        hidden_state = input
+        for _, layer in enumerate(self.layers):
+            hidden_state = layer(hidden_state)
+        return hidden_state
+
+
+class BitEncoder(nn.Module):
+    def __init__(self, config: BitConfig):
+        super().__init__()
+        self.stages = nn.ModuleList([])
+
+        prev_chs = config.embedding_size
+
+        # These needs to stay hardcoded
+        current_stride = 4
+        dilation = 1
+
+        layer_dropouts = [
+            x.tolist() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths)).split(config.depths)
+        ]
+
+        for stage_idx, (current_depth, current_hidden_size, layer_dropout) in enumerate(
+            zip(config.depths, config.hidden_sizes, layer_dropouts)
+        ):
+            # Get the updated hyper params
+            out_channels, stride, dilation = self._get_updated_hyperparameters(
+                stage_idx, current_stride, current_hidden_size, dilation, config
+            )
+
+            stage = BitStage(
+                config,
+                prev_chs,
+                out_channels,
+                stride=stride,
+                dilation=dilation,
+                depth=current_depth,
+                layer_dropout=layer_dropout,
+            )
+
+            prev_chs = out_channels
+            current_stride *= stride
+
+            self.stages.add_module(str(stage_idx), stage)
+
+    def _get_updated_hyperparameters(self, stage_idx, current_stride, current_hidden_size, dilation, config):
+        out_channels = make_div(current_hidden_size * config.width_factor)
+        stride = 1 if stage_idx == 0 else 2
+        if current_stride >= config.output_stride:
+            dilation *= stride
+            stride = 1
+        return out_channels, stride, dilation
+
+    def forward(
+        self, hidden_state: Tensor, output_hidden_states: bool = False, return_dict: bool = True
+    ) -> BaseModelOutputWithNoAttention:
+        hidden_states = () if output_hidden_states else None
+
+        for stage_module in self.stages:
+            if output_hidden_states:
+                hidden_states = hidden_states + (hidden_state,)
+
+            hidden_state = stage_module(hidden_state)
+
+        if output_hidden_states:
+            hidden_states = hidden_states + (hidden_state,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_state, hidden_states] if v is not None)
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=hidden_state,
+            hidden_states=hidden_states,
+        )
+
+
+class BitPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BitConfig
+    base_model_prefix = "bit"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Conv2d):
+            nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu")
+        elif isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
+            nn.init.constant_(module.weight, 1)
+            nn.init.constant_(module.bias, 0)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, BitModel):
+            module.gradient_checkpointing = value
+
+    @torch.no_grad()
+    def _get_feature_map(self, dummy_image):
+        training = self.training
+        if training:
+            self.eval()
+        feature_map = self(dummy_image).feature_maps[-1]
+        self.train(training)
+        return feature_map
+
+
+BIT_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 ([`BitConfig`]): 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.
+"""
+
+BIT_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoFeatureExtractor`]. See
+            [`AutoFeatureExtractor.__call__`] for details.
+
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare BiT model outputting raw features without any specific head on top.",
+    BIT_START_DOCSTRING,
+)
+class BitModel(BitPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embedder = BitEmbeddings(config)
+
+        self.encoder = BitEncoder(config)
+        self.norm = (
+            BitGroupNormActivation(config, num_channels=config.hidden_sizes[-1])
+            if config.layer_type == "preactivation"
+            else nn.Identity()
+        )
+
+        self.pooler = nn.AdaptiveAvgPool2d((1, 1))
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BIT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        processor_class=_FEAT_EXTRACTOR_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPoolingAndNoAttention,
+        config_class=_CONFIG_FOR_DOC,
+        modality="vision",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self, pixel_values: Tensor, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None
+    ) -> BaseModelOutputWithPoolingAndNoAttention:
+        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
+
+        embedding_output = self.embedder(pixel_values)
+
+        encoder_outputs = self.encoder(
+            embedding_output, output_hidden_states=output_hidden_states, return_dict=return_dict
+        )
+
+        last_hidden_state = encoder_outputs[0]
+
+        last_hidden_state = self.norm(last_hidden_state)
+
+        pooled_output = self.pooler(last_hidden_state)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndNoAttention(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+        )
+
+
+@add_start_docstrings(
+    """
+    BiT Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for
+    ImageNet.
+    """,
+    BIT_START_DOCSTRING,
+)
+class BitForImageClassification(BitPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.bit = BitModel(config)
+        # classification head
+        self.classifier = nn.Sequential(
+            nn.Flatten(),
+            nn.Linear(config.hidden_sizes[-1], config.num_labels) if config.num_labels > 0 else nn.Identity(),
+        )
+        # initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BIT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        processor_class=_FEAT_EXTRACTOR_FOR_DOC,
+        checkpoint=_IMAGE_CLASS_CHECKPOINT,
+        output_type=ImageClassifierOutputWithNoAttention,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> ImageClassifierOutputWithNoAttention:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.bit(pixel_values, output_hidden_states=output_hidden_states, return_dict=return_dict)
+
+        pooled_output = outputs.pooler_output if return_dict else outputs[1]
+
+        logits = self.classifier(pooled_output)
+
+        loss = None
+
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[2:]
+            return (loss,) + output if loss is not None else output
+
+        return ImageClassifierOutputWithNoAttention(loss=loss, logits=logits, hidden_states=outputs.hidden_states)
+
+
+@add_start_docstrings(
+    """
+    BiT backbone, to be used with frameworks like DETR and MaskFormer.
+    """,
+    BIT_START_DOCSTRING,
+)
+class BitBackbone(BitPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.stage_names = config.stage_names
+        self.bit = BitModel(config)
+
+        self.out_features = config.out_features
+
+        out_feature_channels = {}
+        out_feature_channels["stem"] = config.embedding_size
+        for idx, stage in enumerate(self.stage_names[1:]):
+            out_feature_channels[stage] = config.hidden_sizes[idx]
+
+        self.out_feature_channels = out_feature_channels
+
+        # initialize weights and apply final processing
+        self.post_init()
+
+    @property
+    def channels(self):
+        return [self.out_feature_channels[name] for name in self.out_features]
+
+    @add_start_docstrings_to_model_forward(BIT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BackboneOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self, pixel_values: Tensor, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None
+    ) -> BackboneOutput:
+        """
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import 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("google/resnetnv2-50")
+        >>> model = AutoBackbone.from_pretrained("google/resnetnv2-50")
+
+        >>> 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
+        )
+
+        outputs = self.bit(pixel_values, output_hidden_states=True, return_dict=True)
+
+        hidden_states = outputs.hidden_states
+
+        feature_maps = ()
+        for idx, stage in enumerate(self.stage_names):
+            if stage in self.out_features:
+                feature_maps += (hidden_states[idx],)
+
+        if not return_dict:
+            output = (feature_maps,)
+            if output_hidden_states:
+                output += (outputs.hidden_states,)
+            return output
+
+        return BackboneOutput(
+            feature_maps=feature_maps,
+            hidden_states=outputs.hidden_states if output_hidden_states else None,
+            attentions=None,
+        )

src/transformers/models/vit_hybrid/__init__.py

@@ -0,0 +1,75 @@
+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available, is_vision_available
+
+
+_import_structure = {"configuration_vit_hybrid": ["VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP", "ViTHybridConfig"]}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_vit_hybrid"] = [
+        "VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "ViTHybridForImageClassification",
+        "ViTHybridModel",
+        "ViTHybridPreTrainedModel",
+    ]
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_vit_hybrid"] = ["ViTHybridImageProcessor"]
+
+
+if TYPE_CHECKING:
+    from .configuration_vit_hybrid import VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP, ViTHybridConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_vit_hybrid import (
+            VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ViTHybridForImageClassification,
+            ViTHybridModel,
+            ViTHybridPreTrainedModel,
+        )
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_vit_hybrid import ViTHybridImageProcessor
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/vit_hybrid/configuration_vit_hybrid.py

@@ -0,0 +1,152 @@
+# coding=utf-8
+# Copyright 2022 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.
+""" ViT Hybrid model configuration"""
+
+import copy
+from typing import Dict
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ..auto.configuration_auto import CONFIG_MAPPING
+from ..bit import BitConfig
+
+
+logger = logging.get_logger(__name__)
+
+VIT_HYBRID_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "google/vit-hybrid-base-bit-384": "https://huggingface.co/vit-hybrid-base-bit-384/resolve/main/config.json",
+    # See all ViT hybrid models at https://huggingface.co/models?filter=vit
+}
+
+
+class ViTHybridConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`ViTHybridModel`]. It is used to instantiate a ViT
+    Hybrid 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 ViT Hybrid
+    [google/vit-hybrid-base-bit-384](https://huggingface.co/google/vit-hybrid-base-bit-384) 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.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        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 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            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-12):
+            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 1):
+            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.
+
+    Example:
+
+    ```python
+    >>> from transformers import ViTHybridConfig, ViTHybridModel
+
+    >>> # Initializing a ViT Hybrid vit-hybrid-base-bit-384 style configuration
+    >>> configuration = ViTHybridConfig()
+
+    >>> # Initializing a model (with random weights) from the vit-hybrid-base-bit-384 style configuration
+    >>> model = ViTHybridModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "vit-hybrid"
+
+    def __init__(
+        self,
+        backbone_config=None,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        image_size=224,
+        patch_size=1,
+        num_channels=3,
+        qkv_bias=True,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        if backbone_config is None:
+            logger.info("`backbone_config` is `None`. Initializing the config with a `BiT` backbone.")
+            backbone_config = {
+                "global_padding": "same",
+                "layer_type": "bottleneck",
+                "depths": [3, 4, 9],
+                "out_features": ["stage3"],
+                "embedding_dynamic_padding": True,
+            }
+
+        if isinstance(backbone_config, dict):
+            if "model_type" in backbone_config:
+                backbone_config_class = CONFIG_MAPPING[backbone_config["model_type"]]
+            else:
+                logger.info(
+                    "`model_type` is not found in `backbone_config`. Use `Bit` as the backbone configuration class."
+                )
+                backbone_config_class = BitConfig
+            backbone_config = backbone_config_class(**backbone_config)
+
+        self.backbone_config = backbone_config
+        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.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
+
+    def to_dict(self) -> Dict[str, any]:
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`]. Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["backbone_config"] = self.backbone_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/vit_hybrid/convert_vit_hybrid_timm_to_pytorch.py

@@ -0,0 +1,283 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ViT hybrid checkpoints from the timm library."""
+
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+from PIL import Image
+
+import requests
+import timm
+from huggingface_hub import hf_hub_download
+from timm.data import resolve_data_config
+from timm.data.transforms_factory import create_transform
+from transformers import (
+    BitConfig,
+    ViTHybridConfig,
+    ViTHybridForImageClassification,
+    ViTHybridImageProcessor,
+    ViTHybridModel,
+)
+from transformers.image_utils import PILImageResampling
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def create_rename_keys(config, base_model=False):
+    rename_keys = []
+
+    # fmt: off
+    # stem:
+    rename_keys.append(("cls_token", "vit.embeddings.cls_token"))
+    rename_keys.append(("pos_embed", "vit.embeddings.position_embeddings"))
+
+    rename_keys.append(("patch_embed.proj.weight", "vit.embeddings.patch_embeddings.projection.weight"))
+    rename_keys.append(("patch_embed.proj.bias", "vit.embeddings.patch_embeddings.projection.bias"))
+
+    # backbone
+    rename_keys.append(("patch_embed.backbone.stem.conv.weight", "vit.embeddings.patch_embeddings.backbone.bit.embedder.convolution.weight"))
+    rename_keys.append(("patch_embed.backbone.stem.norm.weight", "vit.embeddings.patch_embeddings.backbone.bit.embedder.norm.weight"))
+    rename_keys.append(("patch_embed.backbone.stem.norm.bias", "vit.embeddings.patch_embeddings.backbone.bit.embedder.norm.bias"))
+
+    for stage_idx in range(len(config.backbone_config.depths)):
+        for layer_idx in range(config.backbone_config.depths[stage_idx]):
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.conv1.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.conv1.weight"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.norm1.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.norm1.weight"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.norm1.bias", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.norm1.bias"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.conv2.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.conv2.weight"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.norm2.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.norm2.weight"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.norm2.bias", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.norm2.bias"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.conv3.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.conv3.weight"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.norm3.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.norm3.weight"))
+            rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.{layer_idx}.norm3.bias", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.{layer_idx}.norm3.bias"))
+
+        rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.0.downsample.conv.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.0.downsample.conv.weight"))
+        rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.0.downsample.norm.weight", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.0.downsample.norm.weight"))
+        rename_keys.append((f"patch_embed.backbone.stages.{stage_idx}.blocks.0.downsample.norm.bias", f"vit.embeddings.patch_embeddings.backbone.bit.encoder.stages.{stage_idx}.layers.0.downsample.norm.bias"))
+
+    # transformer encoder
+    for i in range(config.num_hidden_layers):
+        # encoder layers: output projection, 2 feedforward neural networks and 2 layernorms
+        rename_keys.append((f"blocks.{i}.norm1.weight", f"vit.encoder.layer.{i}.layernorm_before.weight"))
+        rename_keys.append((f"blocks.{i}.norm1.bias", f"vit.encoder.layer.{i}.layernorm_before.bias"))
+        rename_keys.append((f"blocks.{i}.attn.proj.weight", f"vit.encoder.layer.{i}.attention.output.dense.weight"))
+        rename_keys.append((f"blocks.{i}.attn.proj.bias", f"vit.encoder.layer.{i}.attention.output.dense.bias"))
+        rename_keys.append((f"blocks.{i}.norm2.weight", f"vit.encoder.layer.{i}.layernorm_after.weight"))
+        rename_keys.append((f"blocks.{i}.norm2.bias", f"vit.encoder.layer.{i}.layernorm_after.bias"))
+        rename_keys.append((f"blocks.{i}.mlp.fc1.weight", f"vit.encoder.layer.{i}.intermediate.dense.weight"))
+        rename_keys.append((f"blocks.{i}.mlp.fc1.bias", f"vit.encoder.layer.{i}.intermediate.dense.bias"))
+        rename_keys.append((f"blocks.{i}.mlp.fc2.weight", f"vit.encoder.layer.{i}.output.dense.weight"))
+        rename_keys.append((f"blocks.{i}.mlp.fc2.bias", f"vit.encoder.layer.{i}.output.dense.bias"))
+
+    if base_model:
+        # layernorm + pooler
+        rename_keys.extend(
+            [
+                ("norm.weight", "layernorm.weight"),
+                ("norm.bias", "layernorm.bias"),
+                ("pre_logits.fc.weight", "pooler.dense.weight"),
+                ("pre_logits.fc.bias", "pooler.dense.bias"),
+            ]
+        )
+
+        # if just the base model, we should remove "vit" from all keys that start with "vit"
+        rename_keys = [(pair[0], pair[1][4:]) if pair[1].startswith("vit") else pair for pair in rename_keys]
+    else:
+        # layernorm + classification head
+        rename_keys.extend(
+            [
+                ("norm.weight", "vit.layernorm.weight"),
+                ("norm.bias", "vit.layernorm.bias"),
+                ("head.weight", "classifier.weight"),
+                ("head.bias", "classifier.bias"),
+            ]
+        )
+    # fmt: on
+
+    return rename_keys
+
+
+# we split up the matrix of each encoder layer into queries, keys and values
+def read_in_q_k_v(state_dict, config, base_model=False):
+    for i in range(config.num_hidden_layers):
+        if base_model:
+            prefix = ""
+        else:
+            prefix = "vit."
+        # 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"{prefix}encoder.layer.{i}.attention.attention.query.weight"] = in_proj_weight[
+            : config.hidden_size, :
+        ]
+        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.query.bias"] = in_proj_bias[: config.hidden_size]
+        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.key.weight"] = in_proj_weight[
+            config.hidden_size : config.hidden_size * 2, :
+        ]
+        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.key.bias"] = in_proj_bias[
+            config.hidden_size : config.hidden_size * 2
+        ]
+        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.value.weight"] = in_proj_weight[
+            -config.hidden_size :, :
+        ]
+        state_dict[f"{prefix}encoder.layer.{i}.attention.attention.value.bias"] = in_proj_bias[-config.hidden_size :]
+
+
+def remove_classification_head_(state_dict):
+    ignore_keys = ["head.weight", "head.bias"]
+    for k in ignore_keys:
+        state_dict.pop(k, None)
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@torch.no_grad()
+def convert_vit_checkpoint(vit_name, pytorch_dump_folder_path, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to our ViT structure.
+    """
+
+    # define default ViT hybrid configuration
+    backbone_config = BitConfig(
+        global_padding="same",
+        layer_type="bottleneck",
+        depths=(3, 4, 9),
+        out_features=["stage3"],
+        embedding_dynamic_padding=True,
+    )
+    config = ViTHybridConfig(backbone_config=backbone_config, image_size=384, num_labels=1000)
+    base_model = False
+
+    # load original model from timm
+    timm_model = timm.create_model(vit_name, pretrained=True)
+    timm_model.eval()
+
+    # load state_dict of original model, remove and rename some keys
+    state_dict = timm_model.state_dict()
+    if base_model:
+        remove_classification_head_(state_dict)
+    rename_keys = create_rename_keys(config, base_model)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+    read_in_q_k_v(state_dict, config, base_model)
+
+    repo_id = "huggingface/label-files"
+    filename = "imagenet-1k-id2label.json"
+    id2label = json.load(open(hf_hub_download(repo_id, filename, repo_type="dataset"), "r"))
+    id2label = {int(k): v for k, v in id2label.items()}
+    config.id2label = id2label
+    config.label2id = {v: k for k, v in id2label.items()}
+
+    # load HuggingFace model
+    if vit_name[-5:] == "in21k":
+        model = ViTHybridModel(config).eval()
+    else:
+        model = ViTHybridForImageClassification(config).eval()
+    model.load_state_dict(state_dict)
+
+    # create image processor
+    transform = create_transform(**resolve_data_config({}, model=timm_model))
+    timm_transforms = transform.transforms
+
+    pillow_resamplings = {
+        "bilinear": PILImageResampling.BILINEAR,
+        "bicubic": PILImageResampling.BICUBIC,
+        "nearest": PILImageResampling.NEAREST,
+    }
+
+    processor = ViTHybridImageProcessor(
+        do_resize=True,
+        size={"shortest_edge": timm_transforms[0].size},
+        resample=pillow_resamplings[timm_transforms[0].interpolation.value],
+        do_center_crop=True,
+        crop_size={"height": timm_transforms[1].size[0], "width": timm_transforms[1].size[1]},
+        do_normalize=True,
+        image_mean=timm_transforms[-1].mean.tolist(),
+        image_std=timm_transforms[-1].std.tolist(),
+    )
+
+    image = prepare_img()
+    timm_pixel_values = transform(image).unsqueeze(0)
+    pixel_values = processor(image, return_tensors="pt").pixel_values
+
+    # verify pixel values
+    assert torch.allclose(timm_pixel_values, pixel_values)
+
+    # verify logits
+    with torch.no_grad():
+        outputs = model(pixel_values)
+        logits = outputs.logits
+
+    print("Predicted class:", logits.argmax(-1).item())
+    if base_model:
+        timm_pooled_output = timm_model.forward_features(pixel_values)
+        assert timm_pooled_output.shape == outputs.pooler_output.shape
+        assert torch.allclose(timm_pooled_output, outputs.pooler_output, atol=1e-3)
+    else:
+        timm_logits = timm_model(pixel_values)
+        assert timm_logits.shape == outputs.logits.shape
+        assert torch.allclose(timm_logits, outputs.logits, 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 {vit_name} to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        print(f"Saving processor to {pytorch_dump_folder_path}")
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        print(f"Pushing model and processor to the hub {vit_name}")
+        model.push_to_hub(f"ybelkada/{vit_name}")
+        processor.push_to_hub(f"ybelkada/{vit_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--vit_name",
+        default="vit_base_r50_s16_384",
+        type=str,
+        help="Name of the hybrid ViT timm 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 to upload the model to the HuggingFace hub."
+    )
+
+    args = parser.parse_args()
+    convert_vit_checkpoint(args.vit_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/vit_hybrid/image_processing_vit_hybrid.py

@@ -0,0 +1,346 @@
+# coding=utf-8
+# Copyright 2022 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.
+"""Image processor class for ViT hybrid."""
+
+from typing import Any, Dict, List, Optional, Union
+
+import numpy as np
+
+from transformers.utils.generic import TensorType
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from ...image_transforms import (
+    center_crop,
+    get_resize_output_image_size,
+    normalize,
+    rescale,
+    resize,
+    to_channel_dimension_format,
+)
+from ...image_utils import ChannelDimension, ImageInput, PILImageResampling, is_batched, to_numpy_array, valid_images
+from ...utils import logging
+from ...utils.import_utils import is_vision_available
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    import PIL
+
+
+# Copied from transformers.models.bit.image_processing_bit.convert_to_rgb
+def convert_to_rgb(image: Union[Any, PIL.Image.Image]) -> Union[Any, PIL.Image.Image]:
+    """
+    Converts `PIL.Image.Image` to RGB format. Images in other formats are returned as is.
+
+    Args:
+        image (`PIL.Image.Image`):
+            The image to convert.
+    """
+    if not isinstance(image, PIL.Image.Image):
+        return image
+
+    return image.convert("RGB")
+
+
+class ViTHybridImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a ViT Hybrid image processor.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by
+            `do_resize` in the `preprocess` method.
+        size (`Dict[str, int]` *optional*, defaults to `{"shortest_edge": 224}`):
+            Size of the image after resizing. The shortest edge of the image is resized to size["shortest_edge"], with
+            the longest edge resized to keep the input aspect ratio. Can be overridden by `size` in the `preprocess`
+            method.
+        resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Can be overridden by `resample` in the `preprocess` method.
+        do_center_crop (`bool`, *optional*, defaults to `True`):
+            Whether to center crop the image to the specified `crop_size`. Can be overridden by `do_center_crop` in the
+            `preprocess` method.
+        crop_size (`Dict[str, int]` *optional*, defaults to 224):
+            Size of the output image after applying `center_crop`. Can be overridden by `crop_size` in the `preprocess`
+            method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
+            the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
+            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`):
+            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 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
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_center_crop: bool = True,
+        crop_size: Dict[str, int] = None,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        **kwargs
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"shortest_edge": 224}
+        size = get_size_dict(size, default_to_square=False)
+        crop_size = crop_size if crop_size is not None else {"height": 224, "width": 224}
+        crop_size = get_size_dict(crop_size, default_to_square=True, param_name="crop_size")
+
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_center_crop = do_center_crop
+        self.crop_size = crop_size
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else [0.48145466, 0.4578275, 0.40821073]
+        self.image_std = image_std if image_std is not None else [0.26862954, 0.26130258, 0.27577711]
+        self.do_convert_rgb = do_convert_rgb
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Resize an image. The shortest edge of the image is resized to size["shortest_edge"], with the longest edge
+        resized to keep the input aspect ratio.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`Dict[str, int]`):
+                Size of the output image.
+            resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+                Resampling filter to use when resiizing the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        size = get_size_dict(size, default_to_square=False)
+        if "shortest_edge" not in size:
+            raise ValueError(f"The `size` parameter must contain the key `shortest_edge`. Got {size.keys()}")
+        output_size = get_resize_output_image_size(image, size=size["shortest_edge"], default_to_square=False)
+        return resize(image, size=output_size, resample=resample, data_format=data_format, **kwargs)
+
+    def center_crop(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Center crop an image. If the image is too small to be cropped to the size given, it will be padded (so the
+        returned result will always be of size `size`).
+
+        Args:
+            image (`np.ndarray`):
+                Image to center crop.
+            size (`Dict[str, int]`):
+                Size of the output image in the form of a dictionary with keys `height` and `width`.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        size = get_size_dict(size)
+        if "height" not in size or "width" not in size:
+            raise ValueError(f"The `size` parameter must contain the keys (height, width). Got {size.keys()}")
+        return center_crop(image, size=(size["height"], size["width"]), data_format=data_format, **kwargs)
+
+    def rescale(
+        self,
+        image: np.ndarray,
+        scale: Union[int, float],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ):
+        """
+        Rescale an image by a scale factor. image = image * scale.
+
+        Args:
+            image (`np.ndarray`):
+                Image to rescale.
+            scale (`int` or `float`):
+                Scale to apply to the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        return rescale(image, scale=scale, data_format=data_format, **kwargs)
+
+    def normalize(
+        self,
+        image: np.ndarray,
+        mean: Union[float, List[float]],
+        std: Union[float, List[float]],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Normalize an image. image = (image - image_mean) / image_std.
+
+        Args:
+            image (`np.ndarray`):
+                Image to normalize.
+            image_mean (`float` or `List[float]`):
+                Image mean.
+            image_std (`float` or `List[float]`):
+                Image standard deviation.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        return normalize(image, mean=mean, std=std, data_format=data_format, **kwargs)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: bool = None,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = None,
+        do_center_crop: bool = None,
+        crop_size: int = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        **kwargs
+    ) -> PIL.Image.Image:
+        """
+        Preprocess an image or batch of images.
+
+        Args:
+            images (`ImageInput`):
+                Image to preprocess.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing. Shortest edge of the image is resized to size["shortest_edge"], with
+                the longest edge resized to keep the input aspect ratio.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_center_crop (`bool`, *optional*, defaults to `self.do_center_crop`):
+                Whether to center crop the image.
+            crop_size (`Dict[str, int]`, *optional*, defaults to `self.crop_size`):
+                Size of the center crop. Only has an effect if `do_center_crop` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: defaults to the channel dimension format of the input image.
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        size = size if size is not None else self.size
+        size = get_size_dict(size, param_name="size", default_to_square=False)
+        resample = resample if resample is not None else self.resample
+        do_center_crop = do_center_crop if do_center_crop is not None else self.do_center_crop
+        crop_size = crop_size if crop_size is not None else self.crop_size
+        crop_size = get_size_dict(crop_size, param_name="crop_size", default_to_square=True)
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        if not is_batched(images):
+            images = [images]
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        if do_resize and size is None:
+            raise ValueError("Size must be specified if do_resize is True.")
+
+        if do_center_crop and crop_size is None:
+            raise ValueError("Crop size must be specified if do_center_crop is True.")
+
+        if do_rescale and rescale_factor is None:
+            raise ValueError("Rescale factor must be specified if do_rescale is True.")
+
+        if do_normalize and (image_mean is None or image_std is None):
+            raise ValueError("Image mean and std must be specified if do_normalize is True.")
+
+        # PIL RGBA images are converted to RGB
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_resize:
+            images = [self.resize(image=image, size=size, resample=resample) for image in images]
+
+        if do_center_crop:
+            images = [self.center_crop(image=image, size=crop_size) for image in images]
+
+        if do_rescale:
+            images = [self.rescale(image=image, scale=rescale_factor) for image in images]
+
+        if do_normalize:
+            images = [self.normalize(image=image, mean=image_mean, std=image_std) for image in images]
+
+        images = [to_channel_dimension_format(image, data_format) for image in images]
+
+        data = {"pixel_values": images}
+        return BatchFeature(data=data, tensor_type=return_tensors)

src/transformers/models/vit_hybrid/modeling_vit_hybrid.py

@@ -0,0 +1,733 @@
+# coding=utf-8
+# Copyright 2022 Google AI, Ross Wightman, 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 ViT Hybrid model."""
+
+
+import collections.abc
+import math
+from typing import Dict, List, Optional, Set, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling, ImageClassifierOutput
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from ..auto import AutoBackbone
+from .configuration_vit_hybrid import ViTHybridConfig
+
+
+logger = logging.get_logger(__name__)
+
+# General docstring
+_CONFIG_FOR_DOC = "ViTHybridConfig"
+_FEAT_EXTRACTOR_FOR_DOC = "AutoFeatureExtractor"
+
+# Base docstring
+_CHECKPOINT_FOR_DOC = "google/vit-hybrid-base-bit-384"
+_EXPECTED_OUTPUT_SHAPE = [1, 197, 768]
+
+# Image classification docstring
+_IMAGE_CLASS_CHECKPOINT = "google/vit-hybrid-base-bit-384"
+_IMAGE_CLASS_EXPECTED_OUTPUT = "tabby, tabby cat"
+
+
+VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "google/vit-hybrid-base-bit-384",
+    # See all ViT hybrid models at https://huggingface.co/models?filter=vit-hybrid
+]
+
+
+class ViTHybridEmbeddings(nn.Module):
+    """
+    Construct the CLS token, position and patch embeddings. Optionally, also the mask token.
+    """
+
+    def __init__(self, config: ViTHybridConfig, use_mask_token: bool = False) -> None:
+        super().__init__()
+
+        self.cls_token = nn.Parameter(
+            nn.init.trunc_normal_(
+                torch.zeros(1, 1, config.hidden_size, dtype=torch.float32), mean=0.0, std=config.initializer_range
+            )
+        )
+        self.mask_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) if use_mask_token else None
+        self.patch_embeddings = ViTHybridPatchEmbeddings(config)
+        num_patches = self.patch_embeddings.num_patches
+        self.position_embeddings = nn.Parameter(
+            nn.init.trunc_normal_(
+                torch.zeros(1, num_patches + 1, config.hidden_size, dtype=torch.float32),
+                mean=0.0,
+                std=config.initializer_range,
+            )
+        )
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def interpolate_pos_encoding(self, embeddings: torch.Tensor, height: int, width: int) -> torch.Tensor:
+        """
+        This method allows to interpolate the pre-trained position encodings, to be able to use the model on higher
+        resolution images.
+
+        Source:
+        https://github.com/facebookresearch/dino/blob/de9ee3df6cf39fac952ab558447af1fa1365362a/vision_transformer.py#L174
+        """
+
+        num_patches = embeddings.shape[1] - 1
+        num_positions = self.position_embeddings.shape[1] - 1
+        if num_patches == num_positions and height == width:
+            return self.position_embeddings
+        class_pos_embed = self.position_embeddings[:, 0]
+        patch_pos_embed = self.position_embeddings[:, 1:]
+        dim = embeddings.shape[-1]
+        height = height // self.config.patch_size
+        width = width // self.config.patch_size
+        # we add a small number to avoid floating point error in the interpolation
+        # see discussion at https://github.com/facebookresearch/dino/issues/8
+        height, width = height + 0.1, width + 0.1
+        patch_pos_embed = patch_pos_embed.reshape(1, int(math.sqrt(num_positions)), int(math.sqrt(num_positions)), dim)
+        patch_pos_embed = patch_pos_embed.permute(0, 3, 1, 2)
+        patch_pos_embed = nn.functional.interpolate(
+            patch_pos_embed,
+            scale_factor=(height / math.sqrt(num_positions), width / math.sqrt(num_positions)),
+            mode="bicubic",
+            align_corners=False,
+        )
+        if int(height) != patch_pos_embed.shape[-2] or int(width) != patch_pos_embed.shape[-1]:
+            raise ValueError(f"Invalid height or width: {height}, {width}")
+        patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+        return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1)
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        bool_masked_pos: Optional[torch.BoolTensor] = None,
+        interpolate_pos_encoding: bool = False,
+    ) -> torch.Tensor:
+        batch_size, num_channels, height, width = pixel_values.shape
+        embeddings = self.patch_embeddings(pixel_values, interpolate_pos_encoding=interpolate_pos_encoding)
+
+        if bool_masked_pos is not None:
+            seq_length = embeddings.shape[1]
+            mask_tokens = self.mask_token.expand(batch_size, seq_length, -1)
+            # replace the masked visual tokens by mask_tokens
+            mask = bool_masked_pos.unsqueeze(-1).type_as(mask_tokens)
+            embeddings = embeddings * (1.0 - mask) + mask_tokens * mask
+
+        # add the [CLS] token to the embedded patch tokens
+        cls_tokens = self.cls_token.expand(batch_size, -1, -1)
+        embeddings = torch.cat((cls_tokens, embeddings), dim=1)
+
+        # add positional encoding to each token
+        if interpolate_pos_encoding:
+            embeddings = embeddings + self.interpolate_pos_encoding(embeddings, height, width)
+        else:
+            embeddings = embeddings + self.position_embeddings
+
+        embeddings = self.dropout(embeddings)
+
+        return embeddings
+
+
+class ViTHybridPatchEmbeddings(nn.Module):
+    """
+    This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial
+    `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a
+    Transformer.
+    """
+
+    def __init__(self, config, feature_size=None):
+        super().__init__()
+        image_size, patch_size = config.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)
+
+        self.backbone = AutoBackbone.from_config(config.backbone_config)
+        if self.backbone.config.model_type != "bit":
+            raise ValueError(f"Backbone model type {self.backbone.model_type} is not supported.")
+        feature_dim = self.backbone.channels[-1]
+
+        if feature_size is None:
+            dummy_image = torch.zeros(1, num_channels, image_size[0], image_size[1])
+            feature_map = self.backbone._get_feature_map(dummy_image)
+            feature_size = feature_map.shape[-2:]
+            feature_dim = feature_map.shape[1]
+        else:
+            feature_size = (
+                feature_size if isinstance(feature_size, collections.abc.Iterable) else (feature_size, feature_size)
+            )
+            feature_dim = self.backbone.channels[-1]
+
+        self.grid_size = (feature_size[0] // patch_size[0], feature_size[1] // patch_size[1])
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+
+        self.projection = nn.Conv2d(feature_dim, hidden_size, kernel_size=patch_size, stride=patch_size)
+
+    def forward(self, pixel_values: torch.Tensor, interpolate_pos_encoding: bool = False) -> torch.Tensor:
+        _, num_channels, height, width = pixel_values.shape
+        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."
+            )
+        if not interpolate_pos_encoding:
+            if height != self.image_size[0] or width != self.image_size[1]:
+                raise ValueError(
+                    f"Input image size ({height}*{width}) doesn't match model"
+                    f" ({self.image_size[0]}*{self.image_size[1]})."
+                )
+
+        features = self.backbone(pixel_values).feature_maps[-1]
+        embeddings = self.projection(features).flatten(2).transpose(1, 2)
+
+        return embeddings
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfAttention with ViT->ViTHybrid
+class ViTHybridSelfAttention(nn.Module):
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size {config.hidden_size,} is not a multiple of the number of attention "
+                f"heads {config.num_attention_heads}."
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self, hidden_states, head_mask: Optional[torch.Tensor] = None, output_attentions: bool = False
+    ) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor]]:
+        mixed_query_layer = self.query(hidden_states)
+
+        key_layer = self.transpose_for_scores(self.key(hidden_states))
+        value_layer = self.transpose_for_scores(self.value(hidden_states))
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->ViTHybrid
+class ViTHybridSelfOutput(nn.Module):
+    """
+    The residual connection is defined in ViTHybridLayer instead of here (as is the case with other models), due to the
+    layernorm applied before each block.
+    """
+
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTAttention with ViT->ViTHybrid
+class ViTHybridAttention(nn.Module):
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        self.attention = ViTHybridSelfAttention(config)
+        self.output = ViTHybridSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads: Set[int]) -> None:
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    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]]:
+        self_outputs = self.attention(hidden_states, head_mask, output_attentions)
+
+        attention_output = self.output(self_outputs[0], hidden_states)
+
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTIntermediate with ViT->ViTHybrid
+class ViTHybridIntermediate(nn.Module):
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTOutput with ViT->ViTHybrid
+class ViTHybridOutput(nn.Module):
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        hidden_states = hidden_states + input_tensor
+
+        return hidden_states
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTLayer with ViT->ViTHybrid
+class ViTHybridLayer(nn.Module):
+    """This corresponds to the Block class in the timm implementation."""
+
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = ViTHybridAttention(config)
+        self.intermediate = ViTHybridIntermediate(config)
+        self.output = ViTHybridOutput(config)
+        self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    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]]:
+        self_attention_outputs = self.attention(
+            self.layernorm_before(hidden_states),  # in ViTHybrid, layernorm is applied before self-attention
+            head_mask,
+            output_attentions=output_attentions,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        # first residual connection
+        hidden_states = attention_output + hidden_states
+
+        # in ViTHybrid, layernorm is also applied after self-attention
+        layer_output = self.layernorm_after(hidden_states)
+        layer_output = self.intermediate(layer_output)
+
+        # second residual connection is done here
+        layer_output = self.output(layer_output, hidden_states)
+
+        outputs = (layer_output,) + outputs
+
+        return outputs
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTEncoder with ViT->ViTHybrid
+class ViTHybridEncoder(nn.Module):
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([ViTHybridLayer(config) for _ in range(config.num_hidden_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,
+        )
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTPreTrainedModel with ViT->ViTHybrid
+class ViTHybridPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = ViTHybridConfig
+    base_model_prefix = "vit"
+    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)
+
+    def _set_gradient_checkpointing(self, module: ViTHybridEncoder, value: bool = False) -> None:
+        if isinstance(module, ViTHybridEncoder):
+            module.gradient_checkpointing = value
+
+
+VIT_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 ([`ViTHybridConfig`]): 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.
+"""
+
+VIT_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`AutoFeatureExtractor`]. See
+            [`AutoFeatureExtractor.__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 ViT Hybrid Model transformer outputting raw hidden-states without any specific head on top.",
+    VIT_START_DOCSTRING,
+)
+# Copied from transformers.models.vit.modeling_vit.ViTModel with ViT->ViTHybrid
+class ViTHybridModel(ViTHybridPreTrainedModel):
+    def __init__(self, config: ViTHybridConfig, add_pooling_layer: bool = True, use_mask_token: bool = False):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = ViTHybridEmbeddings(config, use_mask_token=use_mask_token)
+        self.encoder = ViTHybridEncoder(config)
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.pooler = ViTHybridPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> ViTHybridPatchEmbeddings:
+        return self.embeddings.patch_embeddings
+
+    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(VIT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        processor_class=_FEAT_EXTRACTOR_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=BaseModelOutputWithPooling,
+        config_class=_CONFIG_FOR_DOC,
+        modality="vision",
+        expected_output=_EXPECTED_OUTPUT_SHAPE,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        bool_masked_pos: Optional[torch.BoolTensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        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)
+
+        # TODO: maybe have a cleaner way to cast the input (from `ImageProcessor` side?)
+        expected_dtype = self.embeddings.patch_embeddings.projection.weight.dtype
+        if pixel_values.dtype != expected_dtype:
+            pixel_values = pixel_values.to(expected_dtype)
+
+        embedding_output = self.embeddings(
+            pixel_values, bool_masked_pos=bool_masked_pos, interpolate_pos_encoding=interpolate_pos_encoding
+        )
+
+        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]
+        sequence_output = self.layernorm(sequence_output)
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            head_outputs = (sequence_output, pooled_output) if pooled_output is not None else (sequence_output,)
+            return head_outputs + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+# Copied from transformers.models.vit.modeling_vit.ViTPooler with ViT->ViTHybrid
+class ViTHybridPooler(nn.Module):
+    def __init__(self, config: ViTHybridConfig):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+@add_start_docstrings(
+    """
+    ViT Hybrid Model transformer with an image classification head on top (a linear layer on top of the final hidden
+    state of the [CLS] token) e.g. for ImageNet.
+    """,
+    VIT_START_DOCSTRING,
+)
+# Copied from transformers.models.vit.modeling_vit.ViTForImageClassification with ViT->ViTHybrid
+class ViTHybridForImageClassification(ViTHybridPreTrainedModel):
+    def __init__(self, config: ViTHybridConfig) -> None:
+        super().__init__(config)
+
+        self.num_labels = config.num_labels
+        self.vit = ViTHybridModel(config, add_pooling_layer=False)
+
+        # Classifier head
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels) if config.num_labels > 0 else nn.Identity()
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(VIT_INPUTS_DOCSTRING)
+    @add_code_sample_docstrings(
+        processor_class=_FEAT_EXTRACTOR_FOR_DOC,
+        checkpoint=_IMAGE_CLASS_CHECKPOINT,
+        output_type=ImageClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+        expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT,
+    )
+    def forward(
+        self,
+        pixel_values: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        interpolate_pos_encoding: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[tuple, ImageClassifierOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the image classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.vit(
+            pixel_values,
+            head_mask=head_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            interpolate_pos_encoding=interpolate_pos_encoding,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        logits = self.classifier(sequence_output[:, 0, :])
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(logits, labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return ImageClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/utils/dummy_pt_objects.py

@@ -1019,6 +1019,37 @@ class BioGptPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+BIT_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class BitBackbone(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BitForImageClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BitModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BitPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 
@@ -5811,6 +5842,30 @@ class ViTPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class ViTHybridForImageClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ViTHybridModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class ViTHybridPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 VIT_MAE_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -43,6 +43,13 @@ class BeitImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class BitImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class ChineseCLIPFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 
@@ -421,6 +428,13 @@ class ViTImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class ViTHybridImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class YolosFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/bit/test_modeling_bit.py

@@ -0,0 +1,298 @@
+# coding=utf-8
+# Copyright 2022 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 Bit model. """
+
+
+import inspect
+import unittest
+
+from transformers import BitConfig
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import cached_property, is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import BitBackbone, BitForImageClassification, BitImageProcessor, BitModel
+    from transformers.models.bit.modeling_bit import BIT_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class BitModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=3,
+        image_size=32,
+        num_channels=3,
+        embeddings_size=10,
+        hidden_sizes=[8, 16, 32, 64],
+        depths=[1, 1, 2, 1],
+        is_training=True,
+        use_labels=True,
+        hidden_act="relu",
+        num_labels=3,
+        scope=None,
+        out_features=["stage2", "stage3", "stage4"],
+        num_groups=1,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.num_channels = num_channels
+        self.embeddings_size = embeddings_size
+        self.hidden_sizes = hidden_sizes
+        self.depths = depths
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.hidden_act = hidden_act
+        self.num_labels = num_labels
+        self.scope = scope
+        self.num_stages = len(hidden_sizes)
+        self.out_features = out_features
+        self.num_groups = num_groups
+
+    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.num_labels)
+
+        config = self.get_config()
+
+        return config, pixel_values, labels
+
+    def get_config(self):
+        return BitConfig(
+            num_channels=self.num_channels,
+            embeddings_size=self.embeddings_size,
+            hidden_sizes=self.hidden_sizes,
+            depths=self.depths,
+            hidden_act=self.hidden_act,
+            num_labels=self.num_labels,
+            out_features=self.out_features,
+            num_groups=self.num_groups,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = BitModel(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_sizes[-1], self.image_size // 32, self.image_size // 32),
+        )
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels):
+        config.num_labels = self.num_labels
+        model = BitForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_backbone(self, config, pixel_values, labels):
+        model = BitBackbone(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_sizes[1], 4, 4])
+
+        # verify channels
+        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
+
+    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 BitModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as Bit does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (BitModel, BitForImageClassification, BitBackbone) if is_torch_available() else ()
+
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+    has_attentions = False
+
+    def setUp(self):
+        self.model_tester = BitModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=BitConfig, has_text_modality=False)
+
+    def test_config(self):
+        self.create_and_test_config_common_properties()
+        self.config_tester.create_and_test_config_to_json_string()
+        self.config_tester.create_and_test_config_to_json_file()
+        self.config_tester.create_and_test_config_from_and_save_pretrained()
+        self.config_tester.create_and_test_config_with_num_labels()
+        self.config_tester.check_config_can_be_init_without_params()
+        self.config_tester.check_config_arguments_init()
+
+    def create_and_test_config_common_properties(self):
+        return
+
+    @unittest.skip(reason="Bit does not output attentions")
+    def test_attention_outputs(self):
+        pass
+
+    @unittest.skip(reason="Bit does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Bit does not support input and output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    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_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config=config)
+            for name, module in model.named_modules():
+                if isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
+                    self.assertTrue(
+                        torch.all(module.weight == 1),
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+                    self.assertTrue(
+                        torch.all(module.bias == 0),
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
+    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.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+
+            expected_num_stages = self.model_tester.num_stages
+            self.assertEqual(len(hidden_states), expected_num_stages + 1)
+
+            # Bit's feature maps are of shape (batch_size, num_channels, height, width)
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [self.model_tester.image_size // 4, self.model_tester.image_size // 4],
+            )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        layers_type = ["preactivation", "bottleneck"]
+        for model_class in self.all_model_classes:
+            for layer_type in layers_type:
+                config.layer_type = layer_type
+                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)
+
+    @unittest.skip(reason="Bit does not use feedforward chunking")
+    def test_feed_forward_chunking(self):
+        pass
+
+    def test_for_image_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BIT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BitModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+    return image
+
+
+@require_torch
+@require_vision
+class BitModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_feature_extractor(self):
+        return (
+            BitImageProcessor.from_pretrained(BIT_PRETRAINED_MODEL_ARCHIVE_LIST[0]) if is_vision_available() else None
+        )
+
+    @slow
+    def test_inference_image_classification_head(self):
+        model = BitForImageClassification.from_pretrained(BIT_PRETRAINED_MODEL_ARCHIVE_LIST[0]).to(torch_device)
+
+        feature_extractor = self.default_feature_extractor
+        image = prepare_img()
+        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 1000))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([[-0.6526, -0.5263, -1.4398]]).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))

tests/models/vit_hybrid/test_modeling_vit_hybrid.py

@@ -0,0 +1,231 @@
+# coding=utf-8
+# Copyright 2022 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 ViT Hybrid model. """
+
+
+import inspect
+import unittest
+
+from transformers import ViTHybridConfig
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import cached_property, is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import ViTHybridForImageClassification, ViTHybridImageProcessor, ViTHybridModel
+    from transformers.models.vit_hybrid.modeling_vit_hybrid import VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+class ViTHybridModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=64,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        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
+
+        # in ViT hybrid, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        # the number of patches is based on the feature map of the backbone, which by default uses an output stride
+        # of 32, which means that the feature map has a spatial resolution of 1/32 of the input image size
+        num_patches = (self.image_size // 32) ** 2
+        self.seq_length = num_patches + 1
+
+    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 ViTHybridConfig(
+            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 = ViTHybridModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels):
+        config.num_labels = self.type_sequence_label_size
+        model = ViTHybridForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_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 ViTHybridModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as ViT does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (ViTHybridModel, ViTHybridForImageClassification) if is_torch_available() else ()
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = ViTHybridModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ViTHybridConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="ViT 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_for_image_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_image_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = ViTHybridModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
+    return image
+
+
+@require_torch
+@require_vision
+class ViTModelIntegrationTest(unittest.TestCase):
+    @cached_property
+    def default_feature_extractor(self):
+        return (
+            ViTHybridImageProcessor.from_pretrained(VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST[0])
+            if is_vision_available()
+            else None
+        )
+
+    @slow
+    def test_inference_image_classification_head(self):
+        model = ViTHybridForImageClassification.from_pretrained(VIT_HYBRID_PRETRAINED_MODEL_ARCHIVE_LIST[0]).to(
+            torch_device
+        )
+
+        feature_extractor = self.default_feature_extractor
+        image = prepare_img()
+        inputs = feature_extractor(images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        expected_shape = torch.Size((1, 1000))
+        self.assertEqual(outputs.logits.shape, expected_shape)
+
+        expected_slice = torch.tensor([-1.9090, -0.4993, -0.2389]).to(torch_device)
+
+        self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))

utils/check_repo.py

@@ -672,6 +672,7 @@ SHOULD_HAVE_THEIR_OWN_PAGE = [
     "PyTorchBenchmarkArguments",
     "TensorFlowBenchmark",
     "TensorFlowBenchmarkArguments",
+    "BitBackbone",
     "MaskFormerSwinBackbone",
     "ResNetBackbone",
     "AutoBackbone",