Add UperNet (#20648)

* First draft

* More improvements

* Add convnext backbone

* Add conversion script

* Add more improvements

* Comment out to_dict

* Add to_dict method

* Add default config

* Fix config

* Fix backbone

* Fix backbone some more

* Add docs, auto mapping, tests

* Fix some tests

* Fix more tests

* Fix more tests

* Add conversion script

* Improve conversion script

* Add support for getting reshaped undownsampled hidden states

* Fix forward pass

* Add print statements

* Comment out set_shift_and_window_size

* More improvements

* Correct downsampling layers conversion

* Fix style

* First draft

* Fix conversion script

* Remove config attribute

* Fix more tests

* Update READMEs

* Update ConvNextBackbone

* Fix ConvNext tests

* Align ConvNext with Swin

* Remove files

* Fix index

* Improve docs

* Add output_attentions to model forward

* Add backbone mixin, improve tests

* More improvements

* Update init_weights

* Fix interpolation of logits

* Add UperNetImageProcessor

* Improve image processor

* Fix image processor

* Remove print statements

* Remove script

* Update import

* Add image processor tests

* Remove print statements

* Fix test

* Add integration test

* Add convnext integration test

* Update docstring

* Fix README

* Simplify config

* Apply suggestions

* Improve docs

* Rename class

* Fix test_initialization

* Fix import

* Address review

* Fix confg

* Convert all checkpoints

* Fix default backbone

* Usage same processor as segformer

* Apply suggestions

* Fix init_weights, update conversion scripts

* Improve config

* Use Auto API instead of creating a new image processor

* Fix docs

* Add doctests

* Remove ResNetConfig dependency

* Add always_partition argument

* Fix rebaseé

* Improve docs

* Convert checkpoints
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
Co-authored-by: Niels Rogge <nielsrogge@Nielss-MBP.localdomain>

README.md

@@ -407,6 +407,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
 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.

README_es.md

@@ -407,6 +407,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
 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.

README_hd.md

@@ -380,6 +380,7 @@ conda install -c huggingface transformers
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler 
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (माइक्रोसॉफ्ट रिसर्च से) साथ में दिया गया पेपर [UniSpeech: यूनिफाइड स्पीच रिप्रेजेंटेशन लर्निंग विद लेबलेड एंड अनलेबल्ड डेटा](https:/ /arxiv.org/abs/2101.07597) चेंगई वांग, यू वू, याओ कियान, केनिची कुमातानी, शुजी लियू, फुरु वेई, माइकल ज़ेंग, ज़ुएदोंग हुआंग द्वारा।
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (माइक्रोसॉफ्ट रिसर्च से) कागज के साथ [UNISPEECH-SAT: यूनिवर्सल स्पीच रिप्रेजेंटेशन लर्निंग विद स्पीकर अवेयर प्री-ट्रेनिंग ](https://arxiv.org/abs/2110.05752) सानयुआन चेन, यू वू, चेंग्यी वांग, झेंगयांग चेन, झूओ चेन, शुजी लियू, जियान वू, याओ कियान, फुरु वेई, जिन्यु ली, जियांगज़ान यू द्वारा पोस्ट किया गया।
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (सिंघुआ यूनिवर्सिटी और ननकाई यूनिवर्सिटी से) साथ में पेपर [विजुअल अटेंशन नेटवर्क](https://arxiv.org/ pdf/2202.09741.pdf) मेंग-हाओ गुओ, चेंग-ज़े लू, झेंग-निंग लियू, मिंग-मिंग चेंग, शि-मिन हू द्वारा।
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (मल्टीमीडिया कम्प्यूटिंग ग्रुप, नानजिंग यूनिवर्सिटी से) साथ में पेपर [वीडियोएमएई: मास्क्ड ऑटोएन्कोडर स्व-पर्यवेक्षित वीडियो प्री-ट्रेनिंग के लिए डेटा-कुशल सीखने वाले हैं] (https://arxiv.org/abs/2203.12602) ज़ान टोंग, यिबिंग सॉन्ग, जुए द्वारा वांग, लिमिन वांग द्वारा पोस्ट किया गया।
 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।

README_ja.md

@@ -442,6 +442,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (Google Research から) Yi Tay, Mostafa Dehghani, Vinh Q から公開された研究論文: [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (Microsoft Research から) Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang から公開された研究論文: [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597)
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (Microsoft Research から) Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu から公開された研究論文: [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752)
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (Peking University から) Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun. から公開された研究論文 [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221)
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (Tsinghua University and Nankai University から) Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu から公開された研究論文: [Visual Attention Network](https://arxiv.org/abs/2202.09741)
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (Multimedia Computing Group, Nanjing University から) Zhan Tong, Yibing Song, Jue Wang, Limin Wang から公開された研究論文: [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602)
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (NAVER AI Lab/Kakao Enterprise/Kakao Brain から) Wonjae Kim, Bokyung Son, Ildoo Kim から公開された研究論文: [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334)

README_ko.md

@@ -357,6 +357,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (Google Research 에서) Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzle 의 [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) 논문과 함께 발표했습니다.
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (Microsoft Research 에서) Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang 의 [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) 논문과 함께 발표했습니다.
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (Microsoft Research 에서) Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu 의 [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) 논문과 함께 발표했습니다.
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (Peking University 에서 제공)은 Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.의 [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221)논문과 함께 발표했습니다.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (Tsinghua University and Nankai University 에서) Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu 의 [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) 논문과 함께 발표했습니다.
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (Multimedia Computing Group, Nanjing University 에서) Zhan Tong, Yibing Song, Jue Wang, Limin Wang 의 [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) 논문과 함께 발표했습니다.
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (NAVER AI Lab/Kakao Enterprise/Kakao Brain 에서) Wonjae Kim, Bokyung Son, Ildoo Kim 의 [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) 논문과 함께 발표했습니다.

README_zh-hans.md

@@ -381,6 +381,7 @@ conda install -c huggingface transformers
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (来自 Microsoft Research) 伴随论文 [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) 由 Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang 发布。
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (来自 Microsoft Research) 伴随论文 [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) 由 Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu 发布。
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (来自 Peking University) 伴随论文 [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) 由 Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun 发布。
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (来自 Tsinghua University and Nankai University) 伴随论文 [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) 由 Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu 发布。
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (来自 Multimedia Computing Group, Nanjing University) 伴随论文 [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) 由 Zhan Tong, Yibing Song, Jue Wang, Limin Wang 发布。
 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。

README_zh-hant.md

@@ -393,6 +393,7 @@ conda install -c huggingface transformers
 1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
+1. **[UPerNet](https://huggingface.co/docs/transformers/main/model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
 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.

docs/source/en/_toctree.yml

@@ -450,6 +450,8 @@
         title: Table Transformer
       - local: model_doc/timesformer
         title: TimeSformer
+      - local: model_doc/upernet
+        title: UperNet
       - local: model_doc/van
         title: VAN
       - local: model_doc/videomae

docs/source/en/index.mdx

@@ -194,6 +194,7 @@ The documentation is organized into five sections:
 1. **[UL2](model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
 1. **[UniSpeechSat](model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
+1. **[UPerNet](model_doc/upernet)** (from Peking University) released with the paper [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221) by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun.
 1. **[VAN](model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
 1. **[VideoMAE](model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
 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.
@@ -363,6 +364,7 @@ Flax), PyTorch, and/or TensorFlow.
 |             TrOCR             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           UniSpeech           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         UniSpeechSat          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            UPerNet            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |              VAN              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           VideoMAE            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             ViLT              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/upernet.mdx

+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# UPerNet
+
+## Overview
+
+The UPerNet model was proposed in [Unified Perceptual Parsing for Scene Understanding](https://arxiv.org/abs/1807.10221)
+by Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun. UPerNet is a general framework to effectively segment
+a wide range of concepts from images, leveraging any vision backbone like [ConvNeXt](convnext) or [Swin](swin).
+
+The abstract from the paper is the following:
+
+*Humans recognize the visual world at multiple levels: we effortlessly categorize scenes and detect objects inside, while also identifying the textures and surfaces of the objects along with their different compositional parts. In this paper, we study a new task called Unified Perceptual Parsing, which requires the machine vision systems to recognize as many visual concepts as possible from a given image. A multi-task framework called UPerNet and a training strategy are developed to learn from heterogeneous image annotations. We benchmark our framework on Unified Perceptual Parsing and show that it is able to effectively segment a wide range of concepts from images. The trained networks are further applied to discover visual knowledge in natural scenes.*
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/blob/main/transformers/model_doc/upernet_architecture.jpg"
+alt="drawing" width="600"/>
+
+<small> UPerNet framework. Taken from the <a href="https://arxiv.org/abs/1807.10221">original paper</a>. </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code is based on OpenMMLab's mmsegmentation [here](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/uper_head.py).
+
+## Usage
+
+UPerNet is a general framework for semantic segmentation. It can be used with any vision backbone, like so:
+
+```py
+from transformers import SwinConfig, UperNetConfig, UperNetForSemanticSegmentation
+
+backbone_config = SwinConfig(out_features=["stage1", "stage2", "stage3", "stage4"])
+
+config = UperNetConfig(backbone_config=backbone_config)
+model = UperNetForSemanticSegmentation(config)
+```
+
+To use another vision backbone, like [ConvNeXt](convnext), simply instantiate the model with the appropriate backbone:
+
+```py
+from transformers import ConvNextConfig, UperNetConfig, UperNetForSemanticSegmentation
+
+backbone_config = ConvNextConfig(out_features=["stage1", "stage2", "stage3", "stage4"])
+
+config = UperNetConfig(backbone_config=backbone_config)
+model = UperNetForSemanticSegmentation(config)
+```
+
+Note that this will randomly initialize all the weights of the model.
+
+## UperNetConfig
+
+[[autodoc]] UperNetConfig
+
+## UperNetForSemanticSegmentation
+
+[[autodoc]] UperNetForSemanticSegmentation
+    - forward
\ No newline at end of file

src/transformers/__init__.py

@@ -424,6 +424,7 @@ _import_structure = {
         "UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "UniSpeechSatConfig",
     ],
+    "models.upernet": ["UperNetConfig"],
     "models.van": ["VAN_PRETRAINED_CONFIG_ARCHIVE_MAP", "VanConfig"],
     "models.videomae": ["VIDEOMAE_PRETRAINED_CONFIG_ARCHIVE_MAP", "VideoMAEConfig"],
     "models.vilt": [
@@ -1224,6 +1225,7 @@ else:
     _import_structure["models.convnext"].extend(
         [
             "CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "ConvNextBackbone",
             "ConvNextForImageClassification",
             "ConvNextModel",
             "ConvNextPreTrainedModel",
@@ -2259,6 +2261,12 @@ else:
             "UniSpeechSatPreTrainedModel",
         ]
     )
+    _import_structure["models.upernet"].extend(
+        [
+            "UperNetForSemanticSegmentation",
+            "UperNetPreTrainedModel",
+        ]
+    )
     _import_structure["models.van"].extend(
         [
             "VAN_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3772,6 +3780,7 @@ if TYPE_CHECKING:
     from .models.trocr import TROCR_PRETRAINED_CONFIG_ARCHIVE_MAP, TrOCRConfig, TrOCRProcessor
     from .models.unispeech import UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechConfig
     from .models.unispeech_sat import UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP, UniSpeechSatConfig
+    from .models.upernet import UperNetConfig
     from .models.van import VAN_PRETRAINED_CONFIG_ARCHIVE_MAP, VanConfig
     from .models.videomae import VIDEOMAE_PRETRAINED_CONFIG_ARCHIVE_MAP, VideoMAEConfig
     from .models.vilt import (
@@ -4456,6 +4465,7 @@ if TYPE_CHECKING:
         )
         from .models.convnext import (
             CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ConvNextBackbone,
             ConvNextForImageClassification,
             ConvNextModel,
             ConvNextPreTrainedModel,
@@ -5292,6 +5302,7 @@ if TYPE_CHECKING:
             UniSpeechSatModel,
             UniSpeechSatPreTrainedModel,
         )
+        from .models.upernet import UperNetForSemanticSegmentation, UperNetPreTrainedModel
         from .models.van import (
             VAN_PRETRAINED_MODEL_ARCHIVE_LIST,
             VanForImageClassification,

src/transformers/models/__init__.py

@@ -166,6 +166,7 @@ from . import (
     trocr,
     unispeech,
     unispeech_sat,
+    upernet,
     van,
     videomae,
     vilt,

src/transformers/models/auto/configuration_auto.py

@@ -162,6 +162,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("trocr", "TrOCRConfig"),
         ("unispeech", "UniSpeechConfig"),
         ("unispeech-sat", "UniSpeechSatConfig"),
+        ("upernet", "UperNetConfig"),
         ("van", "VanConfig"),
         ("videomae", "VideoMAEConfig"),
         ("vilt", "ViltConfig"),
@@ -311,6 +312,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("transfo-xl", "TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("unispeech", "UNISPEECH_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("unispeech-sat", "UNISPEECH_SAT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("upernet", "UPERNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("van", "VAN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("videomae", "VIDEOMAE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("vilt", "VILT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -489,6 +491,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("ul2", "UL2"),
         ("unispeech", "UniSpeech"),
         ("unispeech-sat", "UniSpeechSat"),
+        ("upernet", "UPerNet"),
         ("van", "VAN"),
         ("videomae", "VideoMAE"),
         ("vilt", "ViLT"),

src/transformers/models/auto/image_processing_auto.py

@@ -79,6 +79,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("swinv2", "ViTImageProcessor"),
         ("table-transformer", "DetrImageProcessor"),
         ("timesformer", "VideoMAEImageProcessor"),
+        ("upernet", "SegformerImageProcessor"),
         ("van", "ConvNextImageProcessor"),
         ("videomae", "VideoMAEImageProcessor"),
         ("vilt", "ViltImageProcessor"),

src/transformers/models/auto/modeling_auto.py

@@ -438,6 +438,7 @@ MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES = OrderedDict(
         ("mobilenet_v2", "MobileNetV2ForSemanticSegmentation"),
         ("mobilevit", "MobileViTForSemanticSegmentation"),
         ("segformer", "SegformerForSemanticSegmentation"),
+        ("upernet", "UperNetForSemanticSegmentation"),
     ]
 )
 
@@ -891,6 +892,7 @@ MODEL_FOR_BACKBONE_MAPPING_NAMES = OrderedDict(
     [
         # Backbone mapping
         ("bit", "BitBackbone"),
+        ("convnext", "ConvNextBackbone"),
         ("dinat", "DinatBackbone"),
         ("maskformer-swin", "MaskFormerSwinBackbone"),
         ("nat", "NatBackbone"),

src/transformers/models/convnext/__init__.py

@@ -51,6 +51,7 @@ else:
         "ConvNextForImageClassification",
         "ConvNextModel",
         "ConvNextPreTrainedModel",
+        "ConvNextBackbone",
     ]
 
 try:
@@ -85,6 +86,7 @@ if TYPE_CHECKING:
     else:
         from .modeling_convnext import (
             CONVNEXT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            ConvNextBackbone,
             ConvNextForImageClassification,
             ConvNextModel,
             ConvNextPreTrainedModel,

src/transformers/models/convnext/configuration_convnext.py

@@ -64,6 +64,9 @@ class ConvNextConfig(PretrainedConfig):
             The initial value for the layer scale.
         drop_path_rate (`float`, *optional*, defaults to 0.0):
             The drop rate for stochastic depth.
+        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). Will default to the last stage if unset.
 
     Example:
     ```python
@@ -93,6 +96,7 @@ class ConvNextConfig(PretrainedConfig):
         layer_scale_init_value=1e-6,
         drop_path_rate=0.0,
         image_size=224,
+        out_features=None,
         **kwargs
     ):
         super().__init__(**kwargs)
@@ -108,6 +112,16 @@ class ConvNextConfig(PretrainedConfig):
         self.layer_scale_init_value = layer_scale_init_value
         self.drop_path_rate = drop_path_rate
         self.image_size = image_size
+        self.stage_names = ["stem"] + [f"stage{idx}" for idx in range(1, len(self.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
 
 
 class ConvNextOnnxConfig(OnnxConfig):

src/transformers/models/convnext/modeling_convnext.py

@@ -24,12 +24,19 @@ 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
+from ...modeling_utils import BackboneMixin, PreTrainedModel
+from ...utils import (
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
 from .configuration_convnext import ConvNextConfig
 
 
@@ -290,7 +297,7 @@ class ConvNextPreTrainedModel(PreTrainedModel):
             module.weight.data.fill_(1.0)
 
     def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, ConvNextModel):
+        if isinstance(module, ConvNextEncoder):
             module.gradient_checkpointing = value
 
 
@@ -465,3 +472,102 @@ class ConvNextForImageClassification(ConvNextPreTrainedModel):
             logits=logits,
             hidden_states=outputs.hidden_states,
         )
+
+
+@add_start_docstrings(
+    """
+    ConvNeXt backbone, to be used with frameworks like DETR and MaskFormer.
+    """,
+    CONVNEXT_START_DOCSTRING,
+)
+class ConvNextBackbone(ConvNextPreTrainedModel, BackboneMixin):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.stage_names = config.stage_names
+        self.embeddings = ConvNextEmbeddings(config)
+        self.encoder = ConvNextEncoder(config)
+
+        self.out_features = config.out_features if config.out_features is not None else [self.stage_names[-1]]
+
+        out_feature_channels = {}
+        out_feature_channels["stem"] = config.hidden_sizes[0]
+        for idx, stage in enumerate(self.stage_names[1:]):
+            out_feature_channels[stage] = config.hidden_sizes[idx]
+
+        self.out_feature_channels = out_feature_channels
+
+        # Add layer norms to hidden states of out_features
+        hidden_states_norms = dict()
+        for stage, num_channels in zip(self.out_features, self.channels):
+            hidden_states_norms[stage] = ConvNextLayerNorm(num_channels, data_format="channels_first")
+        self.hidden_states_norms = nn.ModuleDict(hidden_states_norms)
+
+        # 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(CONVNEXT_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BackboneOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        pixel_values: torch.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("facebook/convnext-tiny-224")
+        >>> model = AutoBackbone.from_pretrained("facebook/convnext-tiny-224")
+
+        >>> inputs = processor(image, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+
+        embedding_output = self.embeddings(pixel_values)
+
+        outputs = self.encoder(
+            embedding_output,
+            output_hidden_states=True,
+            return_dict=True,
+        )
+
+        hidden_states = outputs.hidden_states
+
+        feature_maps = ()
+        # we skip the stem
+        for idx, (stage, hidden_state) in enumerate(zip(self.stage_names[1:], hidden_states[1:])):
+            if stage in self.out_features:
+                hidden_state = self.hidden_states_norms[stage](hidden_state)
+                feature_maps += (hidden_state,)
+
+        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/donut/modeling_donut_swin.py

@@ -577,8 +577,12 @@ class DonutSwinLayer(nn.Module):
         input_dimensions: Tuple[int, int],
         head_mask: Optional[torch.FloatTensor] = None,
         output_attentions: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if not always_partition:
             self.set_shift_and_window_size(input_dimensions)
+        else:
+            pass
         height, width = input_dimensions
         batch_size, _, channels = hidden_states.size()
         shortcut = hidden_states
@@ -668,13 +672,16 @@ class DonutSwinStage(nn.Module):
         input_dimensions: Tuple[int, int],
         head_mask: Optional[torch.FloatTensor] = None,
         output_attentions: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
     ) -> Tuple[torch.Tensor]:
         height, width = input_dimensions
         for i, layer_module in enumerate(self.blocks):
 
             layer_head_mask = head_mask[i] if head_mask is not None else None
 
-            layer_outputs = layer_module(hidden_states, input_dimensions, layer_head_mask, output_attentions)
+            layer_outputs = layer_module(
+                hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+            )
 
             hidden_states = layer_outputs[0]
 
@@ -725,6 +732,7 @@ class DonutSwinEncoder(nn.Module):
         output_attentions: Optional[bool] = False,
         output_hidden_states: Optional[bool] = False,
         output_hidden_states_before_downsampling: Optional[bool] = False,
+        always_partition: Optional[bool] = False,
         return_dict: Optional[bool] = True,
     ) -> Union[Tuple, DonutSwinEncoderOutput]:
         all_hidden_states = () if output_hidden_states else None
@@ -754,7 +762,9 @@ class DonutSwinEncoder(nn.Module):
                     create_custom_forward(layer_module), hidden_states, input_dimensions, layer_head_mask
                 )
             else:
-                layer_outputs = layer_module(hidden_states, input_dimensions, layer_head_mask, output_attentions)
+                layer_outputs = layer_module(
+                    hidden_states, input_dimensions, layer_head_mask, output_attentions, always_partition
+                )
 
             hidden_states = layer_outputs[0]
             hidden_states_before_downsampling = layer_outputs[1]

src/transformers/models/segformer/image_processing_segformer.py

@@ -23,7 +23,7 @@ from transformers.utils import is_torch_available, is_torch_tensor, is_vision_av
 from transformers.utils.generic import TensorType
 
 from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
-from ...image_transforms import center_crop, normalize, rescale, resize, to_channel_dimension_format
+from ...image_transforms import normalize, rescale, resize, to_channel_dimension_format
 from ...image_utils import (
     IMAGENET_DEFAULT_MEAN,
     IMAGENET_DEFAULT_STD,
@@ -159,30 +159,6 @@ class SegformerImageProcessor(BaseImageProcessor):
             image, size=(size["height"], size["width"]), 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 to `(size["height"], size["width"])`. If the input size is smaller than `crop_size` along
-        any edge, the image is padded with 0's and then center cropped.
-
-        Args:
-            image (`np.ndarray`):
-                Image to center crop.
-            size (`Dict[str, int]`):
-                Size of the output 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)
-        if "height" not in size or "width" not in size:
-            raise ValueError(f"The `size` dictionary must contain the keys `height` and `width`. Got {size.keys()}")
-        return center_crop(image, size=(size["height"], size["width"]), data_format=data_format, **kwargs)
-
     def rescale(
         self,
         image: np.ndarray,