From cdddfbffa1206958c0fabf60a87561b5704727a8 Mon Sep 17 00:00:00 2001 From: Alara Dirik <8944735+alaradirik@users.noreply.github.com> Date: Tue, 14 Mar 2023 12:08:14 +0300 Subject: [PATCH] Add ConvNeXT V2 (#21679) * Add ConvNeXt V2 to transformers * TF model is separated from the PR to fix issues --- README.md | 1 + README_es.md | 1 + README_hd.md | 1 + README_ja.md | 1 + README_ko.md | 1 + README_zh-hans.md | 1 + README_zh-hant.md | 1 + docs/source/de/index.mdx | 1 + docs/source/en/_toctree.yml | 2 + docs/source/en/index.mdx | 2 + docs/source/en/model_doc/convnextv2.mdx | 57 ++ docs/source/en/tasks/image_classification.mdx | 2 +- docs/source/es/index.mdx | 1 + docs/source/es/serialization.mdx | 1 + docs/source/fr/index.mdx | 1 + docs/source/it/index.mdx | 1 + docs/source/ja/index.mdx | 1 + docs/source/ko/index.mdx | 1 + docs/source/pt/index.mdx | 1 + docs/source/pt/serialization.mdx | 1 + docs/source/zh/index.mdx | 1 + src/transformers/__init__.py | 18 + src/transformers/models/__init__.py | 1 + .../models/auto/configuration_auto.py | 3 + .../models/auto/image_processing_auto.py | 1 + src/transformers/models/auto/modeling_auto.py | 3 + .../convnext/image_processing_convnext.py | 2 +- .../models/convnextv2/__init__.py | 73 +++ .../convnextv2/configuration_convnextv2.py | 114 ++++ .../convert_convnextv2_to_pytorch.py | 282 +++++++++ .../models/convnextv2/modeling_convnextv2.py | 593 ++++++++++++++++++ src/transformers/utils/dummy_pt_objects.py | 31 + tests/models/convnextv2/__init__.py | 0 .../convnextv2/test_modeling_convnextv2.py | 348 ++++++++++ utils/check_repo.py | 1 + 35 files changed, 1548 insertions(+), 2 deletions(-) create mode 100644 docs/source/en/model_doc/convnextv2.mdx create mode 100644 src/transformers/models/convnextv2/__init__.py create mode 100644 src/transformers/models/convnextv2/configuration_convnextv2.py create mode 100644 src/transformers/models/convnextv2/convert_convnextv2_to_pytorch.py create mode 100644 src/transformers/models/convnextv2/modeling_convnextv2.py create mode 100644 tests/models/convnextv2/__init__.py create mode 100644 tests/models/convnextv2/test_modeling_convnextv2.py diff --git a/README.md b/README.md index cd866a4ce4..17df3295ca 100644 --- a/README.md +++ b/README.md @@ -308,6 +308,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/README_es.md b/README_es.md index 3027a03cac..2ae2a604a5 100644 --- a/README_es.md +++ b/README_es.md @@ -296,6 +296,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/README_hd.md b/README_hd.md index fc5fdd882c..9dffd0be1b 100644 --- a/README_hd.md +++ b/README_hd.md @@ -268,6 +268,7 @@ conda install -c huggingface transformers 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (माइक्रोसॉफ्ट रिसर्च एशिया से) कागज के साथ [फास्ट ट्रेनिंग कन्वर्जेंस के लिए सशर्त डीईटीआर](https://arxiv. org/abs/2108.06152) डेपू मेंग, ज़ियाओकांग चेन, ज़ेजिया फैन, गैंग ज़ेंग, होउकियांग ली, युहुई युआन, लेई सन, जिंगडोंग वांग द्वारा। 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech से) साथ में कागज [ConvBERT: स्पैन-आधारित डायनेमिक कनवल्शन के साथ BERT में सुधार](https://arxiv .org/abs/2008.02496) जिहांग जियांग, वीहाओ यू, डाकान झोउ, युनपेंग चेन, जियाशी फेंग, शुइचेंग यान द्वारा। 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI से) साथ वाला पेपर [A ConvNet for the 2020s](https://arxiv.org/abs /2201.03545) ज़ुआंग लियू, हेंज़ी माओ, चाओ-युआन वू, क्रिस्टोफ़ फीचटेनहोफ़र, ट्रेवर डेरेल, सैनिंग ज़ी द्वारा। +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (सिंघुआ यूनिवर्सिटी से) साथ में पेपर [सीपीएम: ए लार्ज-स्केल जेनेरेटिव चाइनीज प्री-ट्रेंड लैंग्वेज मॉडल](https : //arxiv.org/abs/2012.00413) झेंग्यान झांग, जू हान, हाओ झोउ, पेई के, युक्सियन गु, डेमिंग ये, युजिया किन, युशेंग सु, हाओझे जी, जियान गुआन, फैंचाओ क्यूई, ज़ियाओझी वांग, यानान झेंग द्वारा , गुओयांग ज़ेंग, हुआनकी काओ, शेंगकी चेन, डाइक्सुआन ली, ज़ेनबो सन, ज़ियुआन लियू, मिनली हुआंग, वेंटाओ हान, जी तांग, जुआनज़ी ली, ज़ियाओयान झू, माओसोंग सन। 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (सेल्सफोर्स से) साथ में पेपर [CTRL: ए कंडिशनल ट्रांसफॉर्मर लैंग्वेज मॉडल फॉर कंट्रोलेबल जेनरेशन](https://arxiv.org/abs/1909.05858) नीतीश शिरीष केसकर*, ब्रायन मैककैन*, लव आर. वार्ष्णेय, कैमिंग जिओंग और रिचर्ड द्वारा सोचर द्वारा जारी किया गया। 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (Microsoft से) साथ में दिया गया पेपर [CvT: इंट्रोड्यूसिंग कनवॉल्यूशन टू विजन ट्रांसफॉर्मर्स](https://arxiv.org/ एब्स/2103.15808) हैपिंग वू, बिन जिओ, नोएल कोडेला, मेंगचेन लियू, जियांग दाई, लू युआन, लेई झांग द्वारा। diff --git a/README_ja.md b/README_ja.md index 3fae87a294..d08684c036 100644 --- a/README_ja.md +++ b/README_ja.md @@ -330,6 +330,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (Microsoft Research Asia から) Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang から公開された研究論文: [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech から) Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan から公開された研究論文: [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI から) Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie から公開された研究論文: [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (Tsinghua University から) Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun から公開された研究論文: [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (Salesforce から) Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher から公開された研究論文: [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (Microsoft から) Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang から公開された研究論文: [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) diff --git a/README_ko.md b/README_ko.md index c19fabca5c..d68a9b54e2 100644 --- a/README_ko.md +++ b/README_ko.md @@ -245,6 +245,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (Microsoft Research Asia 에서) Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang 의 [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 논문과 함께 발표했습니다. 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech 에서) Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 의 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 논문과 함께 발표했습니다. 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI 에서) Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 의 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 논문과 함께 발표했습니다. +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (Tsinghua University 에서) Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun 의 [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 논문과 함께 발표했습니다. 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (Salesforce 에서) Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher 의 [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) 논문과 함께 발표했습니다. 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (Microsoft 에서) Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang 의 [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) 논문과 함께 발표했습니다. diff --git a/README_zh-hans.md b/README_zh-hans.md index f077fbca40..2e8f9464e0 100644 --- a/README_zh-hans.md +++ b/README_zh-hans.md @@ -269,6 +269,7 @@ conda install -c huggingface transformers 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (来自 Microsoft Research Asia) 伴随论文 [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 由 Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang 发布。 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (来自 YituTech) 伴随论文 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 由 Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 发布。 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (来自 Facebook AI) 伴随论文 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 由 Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 发布。 +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (来自 Tsinghua University) 伴随论文 [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 由 Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun 发布。 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (来自 Salesforce) 伴随论文 [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) 由 Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher 发布。 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (来自 Microsoft) 伴随论文 [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) 由 Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang 发布。 diff --git a/README_zh-hant.md b/README_zh-hant.md index cc19893bc9..7fa3f4a6c3 100644 --- a/README_zh-hant.md +++ b/README_zh-hant.md @@ -281,6 +281,7 @@ conda install -c huggingface transformers 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](https://huggingface.co/docs/transformers/main/model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/docs/source/de/index.mdx b/docs/source/de/index.mdx index 0e2b3fb687..c1340820b5 100644 --- a/docs/source/de/index.mdx +++ b/docs/source/de/index.mdx @@ -73,6 +73,7 @@ Die Bibliothek enthält derzeit JAX-, PyTorch- und TensorFlow-Implementierungen, 1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong. 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index eddfc48d5c..947a409782 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -430,6 +430,8 @@ title: Conditional DETR - local: model_doc/convnext title: ConvNeXT + - local: model_doc/convnextv2 + title: ConvNeXTV2 - local: model_doc/cvt title: CvT - local: model_doc/deformable_detr diff --git a/docs/source/en/index.mdx b/docs/source/en/index.mdx index 36964e924a..1f4f5a0d43 100644 --- a/docs/source/en/index.mdx +++ b/docs/source/en/index.mdx @@ -82,6 +82,7 @@ The documentation is organized into five sections: 1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. @@ -278,6 +279,7 @@ Flax), PyTorch, and/or TensorFlow. | Conditional DETR | ❌ | ❌ | ✅ | ❌ | ❌ | | ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ | | ConvNeXT | ❌ | ❌ | ✅ | ✅ | ❌ | +| ConvNeXTV2 | ❌ | ❌ | ✅ | ❌ | ❌ | | CTRL | ✅ | ❌ | ✅ | ✅ | ❌ | | CvT | ❌ | ❌ | ✅ | ✅ | ❌ | | Data2VecAudio | ❌ | ❌ | ✅ | ❌ | ❌ | diff --git a/docs/source/en/model_doc/convnextv2.mdx b/docs/source/en/model_doc/convnextv2.mdx new file mode 100644 index 0000000000..6572854bfa --- /dev/null +++ b/docs/source/en/model_doc/convnextv2.mdx @@ -0,0 +1,57 @@ + + +# ConvNeXt V2 + +## Overview + +The ConvNeXt V2 model was proposed in [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. +ConvNeXt V2 is a pure convolutional model (ConvNet), inspired by the design of Vision Transformers, and a successor of [ConvNeXT](convnext). + +The abstract from the paper is the following: + +*Driven by improved architectures and better representation learning frameworks, the field of visual recognition has enjoyed rapid modernization and performance boost in the early 2020s. For example, modern ConvNets, represented by ConvNeXt, have demonstrated strong performance in various scenarios. While these models were originally designed for supervised learning with ImageNet labels, they can also potentially benefit from self-supervised learning techniques such as masked autoencoders (MAE). However, we found that simply combining these two approaches leads to subpar performance. In this paper, we propose a fully convolutional masked autoencoder framework and a new Global Response Normalization (GRN) layer that can be added to the ConvNeXt architecture to enhance inter-channel feature competition. This co-design of self-supervised learning techniques and architectural improvement results in a new model family called ConvNeXt V2, which significantly improves the performance of pure ConvNets on various recognition benchmarks, including ImageNet classification, COCO detection, and ADE20K segmentation. We also provide pre-trained ConvNeXt V2 models of various sizes, ranging from an efficient 3.7M-parameter Atto model with 76.7% top-1 accuracy on ImageNet, to a 650M Huge model that achieves a state-of-the-art 88.9% accuracy using only public training data.* + +Tips: + +- See the code examples below each model regarding usage. + + + + ConvNeXt V2 architecture. Taken from the original paper. + +This model was contributed by [adirik](https://huggingface.co/adirik). The original code can be found [here](https://github.com/facebookresearch/ConvNeXt-V2). + +## Resources + +A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with ConvNeXt V2. + + + +- [`ConvNextV2ForImageClassification`] 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. + +## ConvNextV2Config + +[[autodoc]] ConvNextV2Config + +## ConvNextV2Model + +[[autodoc]] ConvNextV2Model + - forward + +## ConvNextV2ForImageClassification + +[[autodoc]] ConvNextV2ForImageClassification + - forward \ No newline at end of file diff --git a/docs/source/en/tasks/image_classification.mdx b/docs/source/en/tasks/image_classification.mdx index 4559b7d1af..c3faa6145e 100644 --- a/docs/source/en/tasks/image_classification.mdx +++ b/docs/source/en/tasks/image_classification.mdx @@ -30,7 +30,7 @@ The task illustrated in this tutorial is supported by the following model archit -[BEiT](../model_doc/beit), [BiT](../model_doc/bit), [ConvNeXT](../model_doc/convnext), [CvT](../model_doc/cvt), [Data2VecVision](../model_doc/data2vec-vision), [DeiT](../model_doc/deit), [DiNAT](../model_doc/dinat), [EfficientFormer](../model_doc/efficientformer), [EfficientNet](../model_doc/efficientnet), [ImageGPT](../model_doc/imagegpt), [LeViT](../model_doc/levit), [MobileNetV1](../model_doc/mobilenet_v1), [MobileNetV2](../model_doc/mobilenet_v2), [MobileViT](../model_doc/mobilevit), [NAT](../model_doc/nat), [Perceiver](../model_doc/perceiver), [PoolFormer](../model_doc/poolformer), [RegNet](../model_doc/regnet), [ResNet](../model_doc/resnet), [SegFormer](../model_doc/segformer), [Swin Transformer](../model_doc/swin), [Swin Transformer V2](../model_doc/swinv2), [VAN](../model_doc/van), [ViT](../model_doc/vit), [ViT Hybrid](../model_doc/vit_hybrid), [ViTMSN](../model_doc/vit_msn) +[BEiT](../model_doc/beit), [BiT](../model_doc/bit), [ConvNeXT](../model_doc/convnext), [ConvNeXTV2](../model_doc/convnextv2), [CvT](../model_doc/cvt), [Data2VecVision](../model_doc/data2vec-vision), [DeiT](../model_doc/deit), [DiNAT](../model_doc/dinat), [EfficientFormer](../model_doc/efficientformer), [EfficientNet](../model_doc/efficientnet), [ImageGPT](../model_doc/imagegpt), [LeViT](../model_doc/levit), [MobileNetV1](../model_doc/mobilenet_v1), [MobileNetV2](../model_doc/mobilenet_v2), [MobileViT](../model_doc/mobilevit), [NAT](../model_doc/nat), [Perceiver](../model_doc/perceiver), [PoolFormer](../model_doc/poolformer), [RegNet](../model_doc/regnet), [ResNet](../model_doc/resnet), [SegFormer](../model_doc/segformer), [Swin Transformer](../model_doc/swin), [Swin Transformer V2](../model_doc/swinv2), [VAN](../model_doc/van), [ViT](../model_doc/vit), [ViT Hybrid](../model_doc/vit_hybrid), [ViTMSN](../model_doc/vit_msn) diff --git a/docs/source/es/index.mdx b/docs/source/es/index.mdx index d9a7a26d24..baedf45a0f 100644 --- a/docs/source/es/index.mdx +++ b/docs/source/es/index.mdx @@ -63,6 +63,7 @@ La biblioteca actualmente contiene implementaciones de JAX, PyTorch y TensorFlow 1. **[CamemBERT](model_doc/camembert)** (de Inria/Facebook/Sorbonne) publicado con el paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) por Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah y Benoît Sagot. 1. **[CANINE](model_doc/canine)** (de Google Research) publicado con el paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) por Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting. 1. **[ConvNeXT](model_doc/convnext)** (de Facebook AI) publicado con el paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) por Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (de Facebook AI) publicado con el paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) por Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CLIP](model_doc/clip)** (de OpenAI) publicado con el paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) por Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever. 1. **[ConvBERT](model_doc/convbert)** (de YituTech) publicado con el paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) por Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[CPM](model_doc/cpm)** (de Universidad de Tsinghua) publicado con el paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) por Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. diff --git a/docs/source/es/serialization.mdx b/docs/source/es/serialization.mdx index 4c42fd5d83..2815734bfa 100644 --- a/docs/source/es/serialization.mdx +++ b/docs/source/es/serialization.mdx @@ -59,6 +59,7 @@ Las configuraciones a la medida incluyen las siguientes arquitecturas: - CodeGen - ConvBERT - ConvNeXT +- ConvNeXTV2 - Data2VecText - Data2VecVision - DeBERTa diff --git a/docs/source/fr/index.mdx b/docs/source/fr/index.mdx index a2caf6309f..8184d1ced8 100644 --- a/docs/source/fr/index.mdx +++ b/docs/source/fr/index.mdx @@ -80,6 +80,7 @@ La documentation est organisée en 5 parties: 1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/docs/source/it/index.mdx b/docs/source/it/index.mdx index e99c64fd1d..a6d474bb6c 100644 --- a/docs/source/it/index.mdx +++ b/docs/source/it/index.mdx @@ -68,6 +68,7 @@ La libreria attualmente contiene implementazioni in JAX, PyTorch e TensorFlow, p 1. **[CamemBERT](model_doc/camembert)** (da Inria/Facebook/Sorbonne) rilasciato con il paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) da Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah e Benoît Sagot. 1. **[CANINE](model_doc/canine)** (da Google Research) rilasciato con il paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) da Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting. 1. **[ConvNeXT](model_doc/convnext)** (da Facebook AI) rilasciato con il paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) da Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (da Facebook AI) rilasciato con il paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) da Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CLIP](model_doc/clip)** (da OpenAI) rilasciato con il paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) da Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever. 1. **[ConvBERT](model_doc/convbert)** (da YituTech) rilasciato con il paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) da Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[CPM](model_doc/cpm)** (dalla Università di Tsinghua) rilasciato con il paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) da Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. diff --git a/docs/source/ja/index.mdx b/docs/source/ja/index.mdx index d98db26346..a644a885d3 100644 --- a/docs/source/ja/index.mdx +++ b/docs/source/ja/index.mdx @@ -78,6 +78,7 @@ specific language governing permissions and limitations under the License. 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (Microsoft Research Asia から) Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang から公開された研究論文: [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech から) Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan から公開された研究論文: [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI から) Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie から公開された研究論文: [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](https://huggingface.co/docs/transformers/model_doc/cpm)** (Tsinghua University から) Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun から公開された研究論文: [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) 1. **[CTRL](https://huggingface.co/docs/transformers/model_doc/ctrl)** (Salesforce から) Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher から公開された研究論文: [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) 1. **[CvT](https://huggingface.co/docs/transformers/model_doc/cvt)** (Microsoft から) Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang から公開された研究論文: [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) diff --git a/docs/source/ko/index.mdx b/docs/source/ko/index.mdx index 8073c2daeb..789aa41a28 100644 --- a/docs/source/ko/index.mdx +++ b/docs/source/ko/index.mdx @@ -72,6 +72,7 @@ specific language governing permissions and limitations under the License. 1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/docs/source/pt/index.mdx b/docs/source/pt/index.mdx index bdbd7385fd..8e74a1feb4 100644 --- a/docs/source/pt/index.mdx +++ b/docs/source/pt/index.mdx @@ -76,6 +76,7 @@ Atualmente a biblioteca contém implementações do PyTorch, TensorFlow e JAX, p 1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot. 1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting. 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever. 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. diff --git a/docs/source/pt/serialization.mdx b/docs/source/pt/serialization.mdx index 2a01640be4..5e95d59512 100644 --- a/docs/source/pt/serialization.mdx +++ b/docs/source/pt/serialization.mdx @@ -60,6 +60,7 @@ As configurações prontas incluem as seguintes arquiteturas: - Conditional DETR - ConvBERT - ConvNeXT +- ConvNeXTV2 - Data2VecText - Data2VecVision - DeBERTa diff --git a/docs/source/zh/index.mdx b/docs/source/zh/index.mdx index 4d69d590c6..33b9181b71 100644 --- a/docs/source/zh/index.mdx +++ b/docs/source/zh/index.mdx @@ -78,6 +78,7 @@ specific language governing permissions and limitations under the License. 1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang. 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan. 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie. +1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie. 1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun. 1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher. 1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang. diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index f66bf314f0..3378935dc7 100644 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -241,6 +241,7 @@ _import_structure = { "models.conditional_detr": ["CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConditionalDetrConfig"], "models.convbert": ["CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvBertConfig", "ConvBertTokenizer"], "models.convnext": ["CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvNextConfig"], + "models.convnextv2": ["CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP", "ConvNextV2Config"], "models.cpm": [], "models.ctrl": ["CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP", "CTRLConfig", "CTRLTokenizer"], "models.cvt": ["CVT_PRETRAINED_CONFIG_ARCHIVE_MAP", "CvtConfig"], @@ -1300,6 +1301,15 @@ else: "ConvNextPreTrainedModel", ] ) + _import_structure["models.convnextv2"].extend( + [ + "CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST", + "ConvNextV2Backbone", + "ConvNextV2ForImageClassification", + "ConvNextV2Model", + "ConvNextV2PreTrainedModel", + ] + ) _import_structure["models.ctrl"].extend( [ "CTRL_PRETRAINED_MODEL_ARCHIVE_LIST", @@ -3851,6 +3861,7 @@ if TYPE_CHECKING: from .models.conditional_detr import CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP, ConditionalDetrConfig from .models.convbert import CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvBertConfig, ConvBertTokenizer from .models.convnext import CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvNextConfig + from .models.convnextv2 import CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvNextV2Config from .models.ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig, CTRLTokenizer from .models.cvt import CVT_PRETRAINED_CONFIG_ARCHIVE_MAP, CvtConfig from .models.data2vec import ( @@ -4777,6 +4788,13 @@ if TYPE_CHECKING: ConvNextModel, ConvNextPreTrainedModel, ) + from .models.convnextv2 import ( + CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST, + ConvNextV2Backbone, + ConvNextV2ForImageClassification, + ConvNextV2Model, + ConvNextV2PreTrainedModel, + ) from .models.ctrl import ( CTRL_PRETRAINED_MODEL_ARCHIVE_LIST, CTRLForSequenceClassification, diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index 29d023fc6a..9951da5538 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -48,6 +48,7 @@ from . import ( conditional_detr, convbert, convnext, + convnextv2, cpm, ctrl, cvt, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index 1ab0b62c79..70ab1bc9b1 100755 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -57,6 +57,7 @@ CONFIG_MAPPING_NAMES = OrderedDict( ("conditional_detr", "ConditionalDetrConfig"), ("convbert", "ConvBertConfig"), ("convnext", "ConvNextConfig"), + ("convnextv2", "ConvNextV2Config"), ("ctrl", "CTRLConfig"), ("cvt", "CvtConfig"), ("data2vec-audio", "Data2VecAudioConfig"), @@ -236,6 +237,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict( ("conditional_detr", "CONDITIONAL_DETR_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("convbert", "CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("convnext", "CONVNEXT_PRETRAINED_CONFIG_ARCHIVE_MAP"), + ("convnextv2", "CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("ctrl", "CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("cvt", "CVT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("data2vec-audio", "DATA2VEC_AUDIO_PRETRAINED_CONFIG_ARCHIVE_MAP"), @@ -405,6 +407,7 @@ MODEL_NAMES_MAPPING = OrderedDict( ("conditional_detr", "Conditional DETR"), ("convbert", "ConvBERT"), ("convnext", "ConvNeXT"), + ("convnextv2", "ConvNeXTV2"), ("cpm", "CPM"), ("ctrl", "CTRL"), ("cvt", "CvT"), diff --git a/src/transformers/models/auto/image_processing_auto.py b/src/transformers/models/auto/image_processing_auto.py index 6274c88f7d..7a042c56c5 100644 --- a/src/transformers/models/auto/image_processing_auto.py +++ b/src/transformers/models/auto/image_processing_auto.py @@ -48,6 +48,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict( ("clipseg", "ViTImageProcessor"), ("conditional_detr", "ConditionalDetrImageProcessor"), ("convnext", "ConvNextImageProcessor"), + ("convnextv2", "ConvNextImageProcessor"), ("cvt", "ConvNextImageProcessor"), ("data2vec-vision", "BeitImageProcessor"), ("deformable_detr", "DeformableDetrImageProcessor"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index f871de8e55..c29965944a 100755 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -56,6 +56,7 @@ MODEL_MAPPING_NAMES = OrderedDict( ("conditional_detr", "ConditionalDetrModel"), ("convbert", "ConvBertModel"), ("convnext", "ConvNextModel"), + ("convnextv2", "ConvNextV2Model"), ("ctrl", "CTRLModel"), ("cvt", "CvtModel"), ("data2vec-audio", "Data2VecAudioModel"), @@ -410,6 +411,7 @@ MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict( ("beit", "BeitForImageClassification"), ("bit", "BitForImageClassification"), ("convnext", "ConvNextForImageClassification"), + ("convnextv2", "ConvNextV2ForImageClassification"), ("cvt", "CvtForImageClassification"), ("data2vec-vision", "Data2VecVisionForImageClassification"), ("deit", ("DeiTForImageClassification", "DeiTForImageClassificationWithTeacher")), @@ -938,6 +940,7 @@ MODEL_FOR_BACKBONE_MAPPING_NAMES = OrderedDict( # Backbone mapping ("bit", "BitBackbone"), ("convnext", "ConvNextBackbone"), + ("convnextv2", "ConvNextV2Backbone"), ("dinat", "DinatBackbone"), ("maskformer-swin", "MaskFormerSwinBackbone"), ("nat", "NatBackbone"), diff --git a/src/transformers/models/convnext/image_processing_convnext.py b/src/transformers/models/convnext/image_processing_convnext.py index 6d1892ef87..5e4d30c5c9 100644 --- a/src/transformers/models/convnext/image_processing_convnext.py +++ b/src/transformers/models/convnext/image_processing_convnext.py @@ -61,7 +61,7 @@ class ConvNextImageProcessor(BaseImageProcessor): be matched to `int(size["shortest_edge"]/crop_pct)`, after which the image is cropped to `(size["shortest_edge"], size["shortest_edge"])`. Only has an effect if `do_resize` is set to `True`. Can be overriden by `size` in the `preprocess` method. - crop_pct (`float` *optional*, defaults to 244 / 256): + crop_pct (`float` *optional*, defaults to 224 / 256): Percentage of the image to crop. Only has an effect if `do_resize` is `True` and size < 384. Can be overriden by `crop_pct` in the `preprocess` method. resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BILINEAR`): diff --git a/src/transformers/models/convnextv2/__init__.py b/src/transformers/models/convnextv2/__init__.py new file mode 100644 index 0000000000..9bfd6b26e0 --- /dev/null +++ b/src/transformers/models/convnextv2/__init__.py @@ -0,0 +1,73 @@ +# 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 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import TYPE_CHECKING + +# rely on isort to merge the imports +from ...utils import ( + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, +) + + +_import_structure = { + "configuration_convnextv2": [ + "CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP", + "ConvNextV2Config", + ] +} + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["modeling_convnextv2"] = [ + "CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST", + "ConvNextV2ForImageClassification", + "ConvNextV2Model", + "ConvNextV2PreTrainedModel", + "ConvNextV2Backbone", + ] + + +if TYPE_CHECKING: + from .configuration_convnextv2 import ( + CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP, + ConvNextV2Config, + ) + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .modeling_convnextv2 import ( + CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST, + ConvNextV2Backbone, + ConvNextV2ForImageClassification, + ConvNextV2Model, + ConvNextV2PreTrainedModel, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure) diff --git a/src/transformers/models/convnextv2/configuration_convnextv2.py b/src/transformers/models/convnextv2/configuration_convnextv2.py new file mode 100644 index 0000000000..4c79d59b10 --- /dev/null +++ b/src/transformers/models/convnextv2/configuration_convnextv2.py @@ -0,0 +1,114 @@ +# coding=utf-8 +# Copyright 2023 Meta Platforms, Inc. 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. +""" ConvNeXTV2 model configuration""" + + +from ...configuration_utils import PretrainedConfig +from ...utils import logging + + +logger = logging.get_logger(__name__) + +CONVNEXTV2_PRETRAINED_CONFIG_ARCHIVE_MAP = { + "facebook/convnextv2-tiny-1k-224": "https://huggingface.co/facebook/convnextv2-tiny-1k-224/resolve/main/config.json", +} + + +class ConvNextV2Config(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`ConvNextV2Model`]. It is used to instantiate an + ConvNeXTV2 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 ConvNeXTV2 + [facebook/convnextv2-tiny-1k-224](https://huggingface.co/facebook/convnextv2-tiny-1k-224) architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + Args: + num_channels (`int`, *optional*, defaults to 3): + The number of input channels. + patch_size (`int`, optional, defaults to 4): + Patch size to use in the patch embedding layer. + num_stages (`int`, optional, defaults to 4): + The number of stages in the model. + hidden_sizes (`List[int]`, *optional*, defaults to `[96, 192, 384, 768]`): + Dimensionality (hidden size) at each stage. + depths (`List[int]`, *optional*, defaults to `[3, 3, 9, 3]`): + Depth (number of blocks) for each stage. + hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`): + The non-linear activation function (function or string) in each block. If string, `"gelu"`, `"relu"`, + `"selu"` and `"gelu_new"` are supported. + 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. + 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 + >>> from transformers import ConvNeXTV2Config, ConvNextV2Model + + >>> # Initializing a ConvNeXTV2 convnextv2-tiny-1k-224 style configuration + >>> configuration = ConvNeXTV2Config() + + >>> # Initializing a model (with random weights) from the convnextv2-tiny-1k-224 style configuration + >>> model = ConvNextV2Model(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + model_type = "convnextv2" + + def __init__( + self, + num_channels=3, + patch_size=4, + num_stages=4, + hidden_sizes=None, + depths=None, + hidden_act="gelu", + initializer_range=0.02, + layer_norm_eps=1e-12, + drop_path_rate=0.0, + image_size=224, + out_features=None, + **kwargs, + ): + super().__init__(**kwargs) + + self.num_channels = num_channels + self.patch_size = patch_size + self.num_stages = num_stages + self.hidden_sizes = [96, 192, 384, 768] if hidden_sizes is None else hidden_sizes + self.depths = [3, 3, 9, 3] if depths is None else depths + self.hidden_act = hidden_act + self.initializer_range = initializer_range + self.layer_norm_eps = layer_norm_eps + 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 diff --git a/src/transformers/models/convnextv2/convert_convnextv2_to_pytorch.py b/src/transformers/models/convnextv2/convert_convnextv2_to_pytorch.py new file mode 100644 index 0000000000..b2a0a52d27 --- /dev/null +++ b/src/transformers/models/convnextv2/convert_convnextv2_to_pytorch.py @@ -0,0 +1,282 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Convert ConvNeXTV2 checkpoints from the original repository. + +URL: https://github.com/facebookresearch/ConvNeXt""" + +import argparse +import json +import os + +import requests +import torch +from huggingface_hub import hf_hub_download +from PIL import Image + +from transformers import ConvNextImageProcessor, ConvNextV2Config, ConvNextV2ForImageClassification +from transformers.image_utils import PILImageResampling +from transformers.utils import logging + + +logging.set_verbosity_info() +logger = logging.get_logger(__name__) + + +def get_convnextv2_config(checkpoint_url): + config = ConvNextV2Config() + + if "atto" in checkpoint_url: + depths = [2, 2, 6, 2] + hidden_sizes = [40, 80, 160, 320] + if "femto" in checkpoint_url: + depths = [2, 2, 6, 2] + hidden_sizes = [48, 96, 192, 384] + if "pico" in checkpoint_url: + depths = [2, 2, 6, 2] + hidden_sizes = [64, 128, 256, 512] + if "nano" in checkpoint_url: + depths = [2, 2, 8, 2] + hidden_sizes = [80, 160, 320, 640] + if "tiny" in checkpoint_url: + depths = [3, 3, 9, 3] + hidden_sizes = [96, 192, 384, 768] + if "base" in checkpoint_url: + depths = [3, 3, 27, 3] + hidden_sizes = [128, 256, 512, 1024] + if "large" in checkpoint_url: + depths = [3, 3, 27, 3] + hidden_sizes = [192, 384, 768, 1536] + if "huge" in checkpoint_url: + depths = [3, 3, 27, 3] + hidden_sizes = [352, 704, 1408, 2816] + + num_labels = 1000 + filename = "imagenet-1k-id2label.json" + expected_shape = (1, 1000) + + repo_id = "huggingface/label-files" + config.num_labels = num_labels + 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()} + config.hidden_sizes = hidden_sizes + config.depths = depths + + return config, expected_shape + + +def rename_key(name): + if "downsample_layers.0.0" in name: + name = name.replace("downsample_layers.0.0", "embeddings.patch_embeddings") + if "downsample_layers.0.1" in name: + name = name.replace("downsample_layers.0.1", "embeddings.norm") # we rename to layernorm later on + if "downsample_layers.1.0" in name: + name = name.replace("downsample_layers.1.0", "stages.1.downsampling_layer.0") + if "downsample_layers.1.1" in name: + name = name.replace("downsample_layers.1.1", "stages.1.downsampling_layer.1") + if "downsample_layers.2.0" in name: + name = name.replace("downsample_layers.2.0", "stages.2.downsampling_layer.0") + if "downsample_layers.2.1" in name: + name = name.replace("downsample_layers.2.1", "stages.2.downsampling_layer.1") + if "downsample_layers.3.0" in name: + name = name.replace("downsample_layers.3.0", "stages.3.downsampling_layer.0") + if "downsample_layers.3.1" in name: + name = name.replace("downsample_layers.3.1", "stages.3.downsampling_layer.1") + if "stages" in name and "downsampling_layer" not in name: + # stages.0.0. for instance should be renamed to stages.0.layers.0. + name = name[: len("stages.0")] + ".layers" + name[len("stages.0") :] + if "stages" in name: + name = name.replace("stages", "encoder.stages") + if "norm" in name: + name = name.replace("norm", "layernorm") + if "head" in name: + name = name.replace("head", "classifier") + + 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 + + +def convert_preprocessor(checkpoint_url): + if "224" in checkpoint_url: + size = 224 + crop_pct = 224 / 256 + elif "384" in checkpoint_url: + size = 384 + crop_pct = None + else: + size = 512 + crop_pct = None + + return ConvNextImageProcessor( + size=size, + crop_pct=crop_pct, + image_mean=[0.485, 0.456, 0.406], + image_std=[0.229, 0.224, 0.225], + resample=PILImageResampling.BICUBIC, + ) + + +@torch.no_grad() +def convert_convnextv2_checkpoint(checkpoint_url, pytorch_dump_folder_path, save_model, push_to_hub): + """ + Copy/paste/tweak model's weights to our ConvNeXTV2 structure. + """ + print("Downloading original model from checkpoint...") + # define ConvNeXTV2 configuration based on URL + config, expected_shape = get_convnextv2_config(checkpoint_url) + # load original state_dict from URL + state_dict = torch.hub.load_state_dict_from_url(checkpoint_url)["model"] + + print("Converting model parameters...") + # rename keys + for key in state_dict.copy().keys(): + val = state_dict.pop(key) + state_dict[rename_key(key)] = val + # add prefix to all keys expect classifier head + for key in state_dict.copy().keys(): + val = state_dict.pop(key) + if not key.startswith("classifier"): + key = "convnextv2." + key + state_dict[key] = val + + # load HuggingFace model + model = ConvNextV2ForImageClassification(config) + model.load_state_dict(state_dict) + model.eval() + + # Check outputs on an image, prepared by ConvNextImageProcessor + preprocessor = convert_preprocessor(checkpoint_url) + inputs = preprocessor(images=prepare_img(), return_tensors="pt") + logits = model(**inputs).logits + + # note: the logits below were obtained without center cropping + if checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_atto_1k_224_ema.pt": + expected_logits = torch.tensor([-0.3930, 0.1747, -0.5246, 0.4177, 0.4295]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_femto_1k_224_ema.pt": + expected_logits = torch.tensor([-0.1727, -0.5341, -0.7818, -0.4745, -0.6566]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_pico_1k_224_ema.pt": + expected_logits = torch.tensor([-0.0333, 0.1563, -0.9137, 0.1054, 0.0381]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_nano_1k_224_ema.pt": + expected_logits = torch.tensor([-0.1744, -0.1555, -0.0713, 0.0950, -0.1431]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_tiny_1k_224_ema.pt": + expected_logits = torch.tensor([0.9996, 0.1966, -0.4386, -0.3472, 0.6661]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_base_1k_224_ema.pt": + expected_logits = torch.tensor([-0.2553, -0.6708, -0.1359, 0.2518, -0.2488]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_large_1k_224_ema.pt": + expected_logits = torch.tensor([-0.0673, -0.5627, -0.3753, -0.2722, 0.0178]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_huge_1k_224_ema.pt": + expected_logits = torch.tensor([-0.6377, -0.7458, -0.2150, 0.1184, -0.0597]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_nano_22k_224_ema.pt": + expected_logits = torch.tensor([1.0799, 0.2322, -0.8860, 1.0219, 0.6231]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_nano_22k_384_ema.pt": + expected_logits = torch.tensor([0.3766, 0.4917, -1.1426, 0.9942, 0.6024]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_tiny_22k_224_ema.pt": + expected_logits = torch.tensor([0.4220, -0.6919, -0.4317, -0.2881, -0.6609]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_tiny_22k_384_ema.pt": + expected_logits = torch.tensor([0.1082, -0.8286, -0.5095, 0.4681, -0.8085]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_base_22k_224_ema.pt": + expected_logits = torch.tensor([-0.2419, -0.6221, 0.2176, -0.0980, -0.7527]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_base_22k_384_ema.pt": + expected_logits = torch.tensor([0.0391, -0.4371, 0.3786, 0.1251, -0.2784]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_large_22k_224_ema.pt": + expected_logits = torch.tensor([-0.0504, 0.5636, -0.1729, -0.6507, -0.3949]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_large_22k_384_ema.pt": + expected_logits = torch.tensor([0.3560, 0.9486, 0.3149, -0.2667, -0.5138]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_huge_22k_384_ema.pt": + expected_logits = torch.tensor([-0.2469, -0.4550, -0.5853, -0.0810, 0.0309]) + elif checkpoint_url == "https://dl.fbaipublicfiles.com/convnext/convnextv2/im22k/convnextv2_huge_22k_512_ema.pt": + expected_logits = torch.tensor([-0.3090, 0.0802, -0.0682, -0.1979, -0.2826]) + else: + raise ValueError(f"Unknown URL: {checkpoint_url}") + + assert torch.allclose(logits[0, :5], expected_logits, atol=1e-3) + assert logits.shape == expected_shape + print("Model outputs match the original results!") + + if save_model: + print("Saving model to local...") + # Create folder to save model + if not os.path.isdir(pytorch_dump_folder_path): + os.mkdir(pytorch_dump_folder_path) + + model.save_pretrained(pytorch_dump_folder_path) + preprocessor.save_pretrained(pytorch_dump_folder_path) + + model_name = "convnextv2" + if "atto" in checkpoint_url: + model_name += "-atto" + if "femto" in checkpoint_url: + model_name += "-femto" + if "pico" in checkpoint_url: + model_name += "-pico" + if "nano" in checkpoint_url: + model_name += "-nano" + elif "tiny" in checkpoint_url: + model_name += "-tiny" + elif "base" in checkpoint_url: + model_name += "-base" + elif "large" in checkpoint_url: + model_name += "-large" + elif "huge" in checkpoint_url: + model_name += "-huge" + if "22k" in checkpoint_url and "1k" not in checkpoint_url: + model_name += "-22k" + elif "22k" in checkpoint_url and "1k" in checkpoint_url: + model_name += "-22k-1k" + elif "1k" in checkpoint_url: + model_name += "-1k" + if "224" in checkpoint_url: + model_name += "-224" + elif "384" in checkpoint_url: + model_name += "-384" + elif "512" in checkpoint_url: + model_name += "-512" + + if push_to_hub: + print(f"Pushing {model_name} to the hub...") + model.push_to_hub(model_name) + preprocessor.push_to_hub(model_name) + + +if __name__ == "__main__": + parser = argparse.ArgumentParser() + # Required parameters + parser.add_argument( + "--checkpoint_url", + default="https://dl.fbaipublicfiles.com/convnext/convnextv2/im1k/convnextv2_atto_1k_224_ema.pt", + type=str, + help="URL of the original ConvNeXTV2 checkpoint you'd like to convert.", + ) + parser.add_argument( + "--pytorch_dump_folder_path", + default="model", + type=str, + help="Path to the output PyTorch model directory.", + ) + parser.add_argument("--save_model", action="store_true", help="Save model to local") + parser.add_argument("--push_to_hub", action="store_true", help="Push model and image preprocessor to the hub") + + args = parser.parse_args() + convert_convnextv2_checkpoint( + args.checkpoint_url, args.pytorch_dump_folder_path, args.save_model, args.push_to_hub + ) diff --git a/src/transformers/models/convnextv2/modeling_convnextv2.py b/src/transformers/models/convnextv2/modeling_convnextv2.py new file mode 100644 index 0000000000..3ee1b34f38 --- /dev/null +++ b/src/transformers/models/convnextv2/modeling_convnextv2.py @@ -0,0 +1,593 @@ +# coding=utf-8 +# Copyright 2023 Meta Platforms, Inc. 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 ConvNextV2 model.""" + + +from typing import Optional, 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 ( + BackboneOutput, + BaseModelOutputWithNoAttention, + BaseModelOutputWithPoolingAndNoAttention, + ImageClassifierOutputWithNoAttention, +) +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_convnextv2 import ConvNextV2Config + + +logger = logging.get_logger(__name__) + +# General docstring +_CONFIG_FOR_DOC = "ConvNextV2Config" + +# Base docstring +_CHECKPOINT_FOR_DOC = "facebook/convnextv2-tiny-1k-224" +_EXPECTED_OUTPUT_SHAPE = [1, 768, 7, 7] + +# Image classification docstring +_IMAGE_CLASS_CHECKPOINT = "facebook/convnextv2-tiny-1k-224" +_IMAGE_CLASS_EXPECTED_OUTPUT = "tabby, tabby cat" + +CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST = [ + "facebook/convnextv2-tiny-1k-224", + # See all ConvNextV2 models at https://huggingface.co/models?filter=convnextv2 +] + + +# Copied from transformers.models.beit.modeling_beit.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->ConvNextV2 +class ConvNextV2DropPath(nn.Module): + """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).""" + + def __init__(self, drop_prob: Optional[float] = None) -> None: + super().__init__() + self.drop_prob = drop_prob + + def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: + return drop_path(hidden_states, self.drop_prob, self.training) + + def extra_repr(self) -> str: + return "p={}".format(self.drop_prob) + + +class ConvNextV2GRN(nn.Module): + """GRN (Global Response Normalization) layer""" + + def __init__(self, dim: int): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim)) + + def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor: + # Compute and normalize global spatial feature maps + global_features = torch.norm(hidden_states, p=2, dim=(1, 2), keepdim=True) + norm_features = global_features / (global_features.mean(dim=-1, keepdim=True) + 1e-6) + hidden_states = self.gamma * (hidden_states * norm_features) + self.beta + hidden_states + + return hidden_states + + +# Copied from transformers.models.convnext.modeling_convnext.ConvNextLayerNorm with ConvNext->ConvNextV2 +class ConvNextV2LayerNorm(nn.Module): + r"""LayerNorm that supports two data formats: channels_last (default) or channels_first. + The ordering of the dimensions in the inputs. channels_last corresponds to inputs with shape (batch_size, height, + width, channels) while channels_first corresponds to inputs with shape (batch_size, channels, height, width). + """ + + def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"): + super().__init__() + self.weight = nn.Parameter(torch.ones(normalized_shape)) + self.bias = nn.Parameter(torch.zeros(normalized_shape)) + self.eps = eps + self.data_format = data_format + if self.data_format not in ["channels_last", "channels_first"]: + raise NotImplementedError(f"Unsupported data format: {self.data_format}") + self.normalized_shape = (normalized_shape,) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if self.data_format == "channels_last": + x = torch.nn.functional.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps) + elif self.data_format == "channels_first": + input_dtype = x.dtype + x = x.float() + u = x.mean(1, keepdim=True) + s = (x - u).pow(2).mean(1, keepdim=True) + x = (x - u) / torch.sqrt(s + self.eps) + x = x.to(dtype=input_dtype) + x = self.weight[:, None, None] * x + self.bias[:, None, None] + return x + + +# Copied from transformers.models.convnext.modeling_convnext.ConvNextEmbeddings with ConvNext->ConvNextV2 +class ConvNextV2Embeddings(nn.Module): + """This class is comparable to (and inspired by) the SwinEmbeddings class + found in src/transformers/models/swin/modeling_swin.py. + """ + + def __init__(self, config): + super().__init__() + self.patch_embeddings = nn.Conv2d( + config.num_channels, config.hidden_sizes[0], kernel_size=config.patch_size, stride=config.patch_size + ) + self.layernorm = ConvNextV2LayerNorm(config.hidden_sizes[0], eps=1e-6, data_format="channels_first") + self.num_channels = config.num_channels + + def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor: + num_channels = pixel_values.shape[1] + if num_channels != self.num_channels: + raise ValueError( + "Make sure that the channel dimension of the pixel values match with the one set in the configuration." + ) + embeddings = self.patch_embeddings(pixel_values) + embeddings = self.layernorm(embeddings) + return embeddings + + +class ConvNextV2Layer(nn.Module): + """This corresponds to the `Block` class in the original implementation. + + There are two equivalent implementations: [DwConv, LayerNorm (channels_first), Conv, GELU,1x1 Conv]; all in (N, C, + H, W) (2) [DwConv, Permute to (N, H, W, C), LayerNorm (channels_last), Linear, GELU, Linear]; Permute back + + The authors used (2) as they find it slightly faster in PyTorch. + + Args: + config ([`ConvNextV2Config`]): Model configuration class. + dim (`int`): Number of input channels. + drop_path (`float`): Stochastic depth rate. Default: 0.0. + """ + + def __init__(self, config, dim, drop_path=0): + super().__init__() + # depthwise conv + self.dwconv = nn.Conv2d(dim, dim, kernel_size=7, padding=3, groups=dim) + self.layernorm = ConvNextV2LayerNorm(dim, eps=1e-6) + # pointwise/1x1 convs, implemented with linear layers + self.pwconv1 = nn.Linear(dim, 4 * dim) + self.act = ACT2FN[config.hidden_act] + self.grn = ConvNextV2GRN(4 * dim) + self.pwconv2 = nn.Linear(4 * dim, dim) + self.drop_path = ConvNextV2DropPath(drop_path) if drop_path > 0.0 else nn.Identity() + + def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor: + input = hidden_states + x = self.dwconv(hidden_states) + # (batch_size, num_channels, height, width) -> (batch_size, height, width, num_channels) + x = x.permute(0, 2, 3, 1) + x = self.layernorm(x) + x = self.pwconv1(x) + x = self.act(x) + x = self.grn(x) + x = self.pwconv2(x) + # (batch_size, height, width, num_channels) -> (batch_size, num_channels, height, width) + x = x.permute(0, 3, 1, 2) + + x = input + self.drop_path(x) + return x + + +# Copied from transformers.models.convnext.modeling_convnext.ConvNextStage with ConvNeXT->ConvNeXTV2, ConvNext->ConvNextV2 +class ConvNextV2Stage(nn.Module): + """ConvNeXTV2 stage, consisting of an optional downsampling layer + multiple residual blocks. + + Args: + config ([`ConvNextV2Config`]): Model configuration class. + in_channels (`int`): Number of input channels. + out_channels (`int`): Number of output channels. + depth (`int`): Number of residual blocks. + drop_path_rates(`List[float]`): Stochastic depth rates for each layer. + """ + + def __init__(self, config, in_channels, out_channels, kernel_size=2, stride=2, depth=2, drop_path_rates=None): + super().__init__() + + if in_channels != out_channels or stride > 1: + self.downsampling_layer = nn.Sequential( + ConvNextV2LayerNorm(in_channels, eps=1e-6, data_format="channels_first"), + nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride), + ) + else: + self.downsampling_layer = nn.Identity() + drop_path_rates = drop_path_rates or [0.0] * depth + self.layers = nn.Sequential( + *[ConvNextV2Layer(config, dim=out_channels, drop_path=drop_path_rates[j]) for j in range(depth)] + ) + + def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor: + hidden_states = self.downsampling_layer(hidden_states) + hidden_states = self.layers(hidden_states) + return hidden_states + + +# Copied from transformers.models.convnext.modeling_convnext.ConvNextEncoder with ConvNext->ConvNextV2 +class ConvNextV2Encoder(nn.Module): + def __init__(self, config): + super().__init__() + self.stages = nn.ModuleList() + drop_path_rates = [ + x.tolist() for x in torch.linspace(0, config.drop_path_rate, sum(config.depths)).split(config.depths) + ] + prev_chs = config.hidden_sizes[0] + for i in range(config.num_stages): + out_chs = config.hidden_sizes[i] + stage = ConvNextV2Stage( + config, + in_channels=prev_chs, + out_channels=out_chs, + stride=2 if i > 0 else 1, + depth=config.depths[i], + drop_path_rates=drop_path_rates[i], + ) + self.stages.append(stage) + prev_chs = out_chs + + def forward( + self, + hidden_states: torch.FloatTensor, + output_hidden_states: Optional[bool] = False, + return_dict: Optional[bool] = True, + ) -> Union[Tuple, BaseModelOutputWithNoAttention]: + all_hidden_states = () if output_hidden_states else None + + for i, layer_module in enumerate(self.stages): + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + hidden_states = layer_module(hidden_states) + + 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] if v is not None) + + return BaseModelOutputWithNoAttention( + last_hidden_state=hidden_states, + hidden_states=all_hidden_states, + ) + + +# Copied from transformers.models.convnext.modeling_convnext.ConvNextPreTrainedModel with ConvNext->ConvNextV2, convnext->convnextv2 +class ConvNextV2PreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + config_class = ConvNextV2Config + base_model_prefix = "convnextv2" + main_input_name = "pixel_values" + supports_gradient_checkpointing = True + + def _init_weights(self, module): + """Initialize the weights""" + if isinstance(module, (nn.Linear, nn.Conv2d)): + # Slightly different from the TF version which uses truncated_normal for initialization + # cf https://github.com/pytorch/pytorch/pull/5617 + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + 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, value=False): + if isinstance(module, ConvNextV2Encoder): + module.gradient_checkpointing = value + + +CONVNEXTV2_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 ([`ConvNextV2Config`]): 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. +""" + +CONVNEXTV2_INPUTS_DOCSTRING = r""" + Args: + pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + Pixel values. Pixel values can be obtained using [`ConvNextImageProcessor`]. See + [`ConvNextImageProcessor.__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 ConvNextV2 model outputting raw features without any specific head on top.", + CONVNEXTV2_START_DOCSTRING, +) +# Copied from transformers.models.convnext.modeling_convnext.ConvNextModel with CONVNEXT->CONVNEXTV2, ConvNext->ConvNextV2 +class ConvNextV2Model(ConvNextV2PreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.config = config + + self.embeddings = ConvNextV2Embeddings(config) + self.encoder = ConvNextV2Encoder(config) + + # final layernorm layer + self.layernorm = nn.LayerNorm(config.hidden_sizes[-1], eps=config.layer_norm_eps) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(CONVNEXTV2_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=BaseModelOutputWithPoolingAndNoAttention, + config_class=_CONFIG_FOR_DOC, + modality="vision", + expected_output=_EXPECTED_OUTPUT_SHAPE, + ) + def forward( + self, + pixel_values: torch.FloatTensor = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, 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 + + if pixel_values is None: + raise ValueError("You have to specify pixel_values") + + embedding_output = self.embeddings(pixel_values) + + encoder_outputs = self.encoder( + embedding_output, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + last_hidden_state = encoder_outputs[0] + + # global average pooling, (N, C, H, W) -> (N, C) + pooled_output = self.layernorm(last_hidden_state.mean([-2, -1])) + + 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( + """ + ConvNextV2 Model with an image classification head on top (a linear layer on top of the pooled features), e.g. for + ImageNet. + """, + CONVNEXTV2_START_DOCSTRING, +) +# Copied from transformers.models.convnext.modeling_convnext.ConvNextForImageClassification with CONVNEXT->CONVNEXTV2,ConvNext->ConvNextV2,convnext->convnextv2 +class ConvNextV2ForImageClassification(ConvNextV2PreTrainedModel): + def __init__(self, config): + super().__init__(config) + + self.num_labels = config.num_labels + self.convnextv2 = ConvNextV2Model(config) + + # Classifier head + self.classifier = ( + 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(CONVNEXTV2_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_IMAGE_CLASS_CHECKPOINT, + output_type=ImageClassifierOutputWithNoAttention, + config_class=_CONFIG_FOR_DOC, + expected_output=_IMAGE_CLASS_EXPECTED_OUTPUT, + ) + def forward( + self, + pixel_values: torch.FloatTensor = None, + labels: Optional[torch.LongTensor] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, 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 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.convnextv2(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( + """ + ConvNeXT V2 backbone, to be used with frameworks like DETR and MaskFormer. + """, + CONVNEXTV2_START_DOCSTRING, +) +# Copied from transformers.models.convnext.modeling_convnext.ConvNextBackbone with CONVNEXT->CONVNEXTV2,ConvNext->ConvNextV2,facebook/convnext-tiny-224->facebook/convnextv2-tiny-1k-224 +class ConvNextV2Backbone(ConvNextV2PreTrainedModel, BackboneMixin): + def __init__(self, config): + super().__init__(config) + + self.stage_names = config.stage_names + self.embeddings = ConvNextV2Embeddings(config) + self.encoder = ConvNextV2Encoder(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 = {} + for stage, num_channels in zip(self.out_features, self.channels): + hidden_states_norms[stage] = ConvNextV2LayerNorm(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(CONVNEXTV2_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/convnextv2-tiny-1k-224") + >>> model = AutoBackbone.from_pretrained("facebook/convnextv2-tiny-1k-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, + ) diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index 1f5931885f..fc9a8d324a 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -1803,6 +1803,37 @@ class ConvNextPreTrainedModel(metaclass=DummyObject): requires_backends(self, ["torch"]) +CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST = None + + +class ConvNextV2Backbone(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class ConvNextV2ForImageClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class ConvNextV2Model(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class ConvNextV2PreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + CTRL_PRETRAINED_MODEL_ARCHIVE_LIST = None diff --git a/tests/models/convnextv2/__init__.py b/tests/models/convnextv2/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/models/convnextv2/test_modeling_convnextv2.py b/tests/models/convnextv2/test_modeling_convnextv2.py new file mode 100644 index 0000000000..10ae34c22d --- /dev/null +++ b/tests/models/convnextv2/test_modeling_convnextv2.py @@ -0,0 +1,348 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" Testing suite for the PyTorch ConvNextV2 model. """ + + +import inspect +import unittest + +from transformers import ConvNextV2Config +from transformers.models.auto import get_values +from transformers.models.auto.modeling_auto import MODEL_FOR_BACKBONE_MAPPING_NAMES, MODEL_MAPPING_NAMES +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 transformers import ConvNextV2Backbone, ConvNextV2ForImageClassification, ConvNextV2Model + from transformers.models.convnextv2.modeling_convnextv2 import CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST + + +if is_vision_available(): + from PIL import Image + + from transformers import AutoImageProcessor + + +class ConvNextV2ModelTester: + def __init__( + self, + parent, + batch_size=13, + image_size=32, + num_channels=3, + num_stages=4, + hidden_sizes=[10, 20, 30, 40], + depths=[2, 2, 3, 2], + is_training=True, + use_labels=True, + intermediate_size=37, + hidden_act="gelu", + num_labels=10, + initializer_range=0.02, + out_features=["stage2", "stage3", "stage4"], + scope=None, + ): + self.parent = parent + self.batch_size = batch_size + self.image_size = image_size + self.num_channels = num_channels + self.num_stages = num_stages + self.hidden_sizes = hidden_sizes + self.depths = depths + self.is_training = is_training + self.use_labels = use_labels + self.intermediate_size = intermediate_size + self.hidden_act = hidden_act + self.num_labels = num_labels + self.initializer_range = initializer_range + self.out_features = out_features + self.scope = scope + + 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 ConvNextV2Config( + num_channels=self.num_channels, + hidden_sizes=self.hidden_sizes, + depths=self.depths, + num_stages=self.num_stages, + hidden_act=self.hidden_act, + is_decoder=False, + initializer_range=self.initializer_range, + out_features=self.out_features, + num_labels=self.num_labels, + ) + + def create_and_check_model(self, config, pixel_values, labels): + model = ConvNextV2Model(config=config) + model.to(torch_device) + model.eval() + result = model(pixel_values) + # expected last hidden states: B, C, H // 32, W // 32 + 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): + model = ConvNextV2ForImageClassification(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 = ConvNextV2Backbone(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:]) + + # verify backbone works with out_features=None + config.out_features = None + model = ConvNextV2Backbone(config=config) + model.to(torch_device) + model.eval() + result = model(pixel_values) + + # verify feature maps + self.parent.assertEqual(len(result.feature_maps), 1) + self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[-1], 1, 1]) + + # verify channels + self.parent.assertEqual(len(model.channels), 1) + 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 + + def prepare_config_and_inputs_with_labels(self): + config_and_inputs = self.prepare_config_and_inputs() + config, pixel_values, labels = config_and_inputs + inputs_dict = {"pixel_values": pixel_values, "labels": labels} + return config, inputs_dict + + +@require_torch +class ConvNextV2ModelTest(ModelTesterMixin, unittest.TestCase): + """ + Here we also overwrite some of the tests of test_modeling_common.py, as ConvNextV2 does not use input_ids, inputs_embeds, + attention_mask and seq_length. + """ + + all_model_classes = ( + ( + ConvNextV2Model, + ConvNextV2ForImageClassification, + ConvNextV2Backbone, + ) + 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 = ConvNextV2ModelTester(self) + self.config_tester = ConfigTester(self, config_class=ConvNextV2Config, has_text_modality=False, hidden_size=37) + + 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="ConvNextV2 does not use inputs_embeds") + def test_inputs_embeds(self): + pass + + @unittest.skip(reason="ConvNextV2 does not support input and output embeddings") + def test_model_common_attributes(self): + pass + + @unittest.skip(reason="ConvNextV2 does not use feedforward chunking") + def test_feed_forward_chunking(self): + pass + + def test_training(self): + if not self.model_tester.is_training: + return + + for model_class in self.all_model_classes: + config, inputs_dict = self.model_tester.prepare_config_and_inputs_with_labels() + config.return_dict = True + + if model_class.__name__ in [ + *get_values(MODEL_MAPPING_NAMES), + *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES), + ]: + continue + + model = model_class(config) + model.to(torch_device) + model.train() + inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True) + loss = model(**inputs).loss + loss.backward() + + def test_training_gradient_checkpointing(self): + if not self.model_tester.is_training: + return + + for model_class in self.all_model_classes: + config, inputs_dict = self.model_tester.prepare_config_and_inputs_with_labels() + config.use_cache = False + config.return_dict = True + + if ( + model_class.__name__ + in [*get_values(MODEL_MAPPING_NAMES), *get_values(MODEL_FOR_BACKBONE_MAPPING_NAMES)] + or not model_class.supports_gradient_checkpointing + ): + continue + + model = model_class(config) + model.to(torch_device) + model.gradient_checkpointing_enable() + model.train() + inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True) + loss = model(**inputs).loss + loss.backward() + + 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_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) + + # ConvNextV2'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() + + for model_class in self.all_model_classes: + inputs_dict["output_hidden_states"] = True + check_hidden_states_output(inputs_dict, config, model_class) + + # check that output_hidden_states also work using config + del inputs_dict["output_hidden_states"] + config.output_hidden_states = True + + check_hidden_states_output(inputs_dict, config, model_class) + + 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 CONVNEXTV2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]: + model = ConvNextV2Model.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 ConvNextV2ModelIntegrationTest(unittest.TestCase): + @cached_property + def default_image_processor(self): + return AutoImageProcessor.from_pretrained("facebook/convnextv2-tiny-1k-224") if is_vision_available() else None + + @slow + def test_inference_image_classification_head(self): + model = ConvNextV2ForImageClassification.from_pretrained("facebook/convnextv2-tiny-1k-224").to(torch_device) + + preprocessor = self.default_image_processor + image = prepare_img() + inputs = preprocessor(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([-0.3083, -0.3040, -0.4344]).to(torch_device) + self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4)) diff --git a/utils/check_repo.py b/utils/check_repo.py index fca0cd86ef..c59897d832 100644 --- a/utils/check_repo.py +++ b/utils/check_repo.py @@ -820,6 +820,7 @@ SHOULD_HAVE_THEIR_OWN_PAGE = [ "AutoBackbone", "BitBackbone", "ConvNextBackbone", + "ConvNextV2Backbone", "DinatBackbone", "MaskFormerSwinBackbone", "MaskFormerSwinConfig",