Hugging Face's logo

BAAI
/
EVA

Model card Files
xet
Community
2
Yuxin-CV commited on
Commit
c00cbbd
·
1 Parent(s): 3bb4d75

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +185 -0
README.md CHANGED
@@ -1,3 +1,188 @@
1
  ---
2
  license: mit
3
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  license: mit
3
  ---
4
+
5
+ <div align="center">
6
+ <h1>EVA: An Open Billion-Scale Vision Foundation Model </h1>
7
+ <h3><a href="https://arxiv.org/abs/2211.07636">EVA: Exploring the Limits of Masked Visual Representation Learning at Scale</a></h3>
8
+
9
+ [Yuxin Fang](https://bit.ly/YuxinFang_GoogleScholar)<sup>2,1</sup>, [Wen Wang](https://scholar.google.com/citations?user=1ks0R04AAAAJ&hl)<sup>3,1</sup>, [Binhui Xie](https://binhuixie.github.io/)<sup>4,1</sup>, [Quan Sun](https://github.com/Quan-Sun)<sup>1</sup>, [Ledell Wu](https://scholar.google.com/citations?user=-eJHVt8AAAAJ&hl=en)<sup>1</sup>, [Xinggang Wang](https://xinggangw.info/)<sup>2</sup>, [Tiejun Huang](https://scholar.google.com/citations?user=knvEK4AAAAAJ&hl=en)<sup>1</sup>, [Xinlong Wang](https://www.xloong.wang/)<sup>1</sup>, [Yue Cao](http://yue-cao.me/)<sup>1</sup>
10
+
11
+ <sup>1</sup>[BAAI](https://www.baai.ac.cn/english.html), <sup>2</sup>[HUST](http://english.hust.edu.cn/), <sup>3</sup>[ZJU](https://www.zju.edu.cn/english/), <sup>4</sup>[BIT](https://english.bit.edu.cn/)
12
+
13
+
14
+ We launch **EVA**, a vision-centric foundation model to **E**xplore the limits of **V**isual representation at sc**A**le using only publicly accessible data and academic resources. **EVA** is a vanilla ViT pre-trained to reconstruct the masked out image-text aligned vision features (*i.e.*, CLIP features) conditioned on visible image patches. Via this pretext task, we can efficiently scale up EVA to one billion parameters, and sets new records on a broad range of representative vision downstream tasks.
15
+
16
+ ***EVA is the first open-sourced billion-scale vision foundation model that achieves state-of-the-art performance on a broad range of downstream tasks.***
17
+
18
+ </div>
19
+
20
+
21
+ **Table of Contents**
22
+
23
+ * [license: mit](#license-mit)
24
+ * [EVA: An Open Billion-Scale Vision Foundation Model ](#eva-an-open-billion-scale-vision-foundation-model-)
25
+ * [<a href="https://arxiv.org/abs/2211.07636" rel="nofollow">EVA: Exploring the Limits of Masked Visual Representation Learning at Scale</a>](#eva-exploring-the-limits-of-masked-visual-representation-learning-at-scale)
26
+ * [Image Classification](#image-classification)
27
+ * [Summary of EVA's image classification performance](#summary-of-evas-image-classification-performance)
28
+ * [Video Classification](#video-classification)
29
+ * [Object Detection &amp; Instance Segmentation](#object-detection--instance-segmentation)
30
+ * [COCO 2017 (single-scale evaluation on val set)](#coco-2017-single-scale-evaluation-on-val-set)
31
+ * [LVIS v1.0 (single-scale evaluation on val set)](#lvis-v10-single-scale-evaluation-on-val-set)
32
+ * [Semantic Segmentation](#semantic-segmentation)
33
+ * [COCO-Stuff-164K](#coco-stuff-164k)
34
+ * [ADE20K](#ade20k)
35
+ * [EVA-CLIP](#eva-clip)
36
+ * [Citation](#citation)
37
+ * [License](#license)
38
+ * [Contact](#contact)
39
+
40
+
41
+ ## Image Classification
42
+
43
+ We provide **all pre-trained & fine-tuned** EVAs for the community.
44
+ The following table summarizes the basic statistics of MIM pre-trained EVA and image classification EVA.
45
+
46
+ | model name | #param. |pre-training epochs on merged-30M | intermeidate fine-tuning epochs on IN-21K | fine-tuning epochs on IN-1K | IN-1K top-1 acc. |weight |
47
+ |------------|:------:|:------------------:|:------:|:------:|:------:|:------:|
48
+ | `eva_psz14` | 1.0B | 150 | - | - | - | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_psz14.pt) (`2GB`) |
49
+ | `eva_psz14to16` | 1.0B | 150 | - | - | - | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_psz14to16.pt) (`2GB`) |
50
+ | `eva_21k_224px_psz14` | 1.0B | 150 | 60 | - | - | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_21k_224px_psz14.pt) (`2GB`) |
51
+ | `eva_21k_1k_336px_psz14_ema` | 1.0B | 150 | 60 | 10 | **89.6** | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_21k_1k_336px_psz14_ema_89p6.pt) (`4GB`) |
52
+ | `eva_21k_1k_560px_psz14_ema` | 1.0B | 150 | 60 | 15 | **89.7** | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_21k_1k_560px_psz14_ema_89p7.pt) (`4GB`) |
53
+
54
+ - `eva_psz14to16` model interpolates the kernel size of `patch_embed` from `14x14` to `16x16`. This is useful for object detection, instance segmentation & semantic segmentation, *etc*. See [`interpolate_patch_14to16.py`](interpolate_patch_14to16.py) for implementation details.
55
+ - For MIM pre-trained EVA and EVA-CLIP, we use `deepspeed` `fp16` format. IN-1K fine-tuned EVA weights are larger (`4GB` *v.s.* `2GB`) because ema updates models with `fp32` format. The weights of other downstream tasks are also with `fp32` format.
56
+
57
+ </div>
58
+
59
+
60
+ ### Summary of EVA's image classification performance
61
+
62
+ <div align="center">
63
+
64
+ | model | [IN-1K](https://github.com/rwightman/pytorch-image-models/blob/main/results/results-imagenet.csv) | [IN-V2](https://github.com/rwightman/pytorch-image-models/blob/main/results/results-imagenetv2-matched-frequency.csv) | [IN-ReaL](https://github.com/rwightman/pytorch-image-models/blob/main/results/results-imagenet-real.csv) | [IN-Adv.](https://github.com/rwightman/pytorch-image-models/blob/main/results/results-imagenet-a.csv) | [IN-Ren.](https://github.com/rwightman/pytorch-image-models/blob/main/results/results-imagenet-r.csv) | [IN-Ske.](https://github.com/rwightman/pytorch-image-models/blob/main/results/results-imagenet-r.csv) | ObjectNet |
65
+ |:------------:|:------------------:|:------:|:------:| :------:|:------:|:------:|:------:|
66
+ | EVA | 89.6 | 81.6 | 90.8 | 86.2 | 88.3 | 67.7 | 60.9 |
67
+
68
+ </div>
69
+
70
+
71
+ ## Video Classification
72
+
73
+ <div align="center">
74
+
75
+ | dataset | model name | init. weight | acc@1 | config | weight | logs |
76
+ |:-----------:|:------------------:|:------------------------------------------------------------------------------------:|:--------:|:--------------------------------------------------:|:----------------------------------------------------------------------------------------:|:----------------------------------------:|
77
+ | Kinetics722 | `eva_video_k722` | [`eva_psz14`](https://huggingface.co/BAAI/EVA/blob/main/eva_psz14.pt) | - | [config](configs/kinetics722_intermediate_ft.yaml) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k722.pth) (`4.8GB`) | [ft_k722](../logs/video/ft_k722_log.txt) |
78
+ | Kinetics400 | `eva_video_k400` | [`eva_video_k722`](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k722.pth) | **89.7** | [config](configs/kinetics400_ft.yaml) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k400.pth) (`4.8GB`) | [ft_k400](../logs/video/ft_k400_log.txt) |
79
+ | Kinetics600 | `eva_video_k600` | [`eva_video_k722`](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k722.pth) | **89.8** | [config](configs/kinetics600_ft.yaml) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k600.pth) (`4.8GB`) | [ft_k600](../logs/video/ft_k600_log.txt) |
80
+ | Kinetics700 | `eva_video_k700` | [`eva_video_k722`](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k722.pth) | **82.9** | [config](configs/kinetics700_ft.yaml) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_video_k700.pth) (`4.8GB`) | [ft_k700](../logs/video/ft_k700_log.txt) |
81
+
82
+ </div>
83
+
84
+
85
+ ## Object Detection & Instance Segmentation
86
+
87
+ <div align="center">
88
+
89
+ | model name | #param. | pre-training interations on Objects365 | weight |
90
+ |------------|:-------:|:--------------------------------------:|:-----------------------------------------------------------------------------:|
91
+ | `eva_o365` | 1.1B | 380k | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_o365.pth) (`4GB`) |
92
+
93
+ </div>
94
+
95
+
96
+ ### COCO 2017 (single-scale evaluation on `val` set)
97
+
98
+ <div align="center">
99
+
100
+ | init. model weight | batch size | iter | AP box | AP mask | config | model weight |
101
+ | :---: | :---: |:----:|:--------:|:---------------:|:---------------------------------------:|:--------------------------------------------------------------------:|
102
+ | [`eva_o365`](https://huggingface.co/BAAI/EVA/blob/main/eva_o365.pth) | 64 | 35k | **64.2** | **53.9** | [config](projects/ViTDet/configs/COCO/cascade_mask_rcnn_vitdet_eva.py) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_coco_det.pth) (`4GB`) |
103
+ | [`eva_o365`](https://huggingface.co/BAAI/EVA/blob/main/eva_o365.pth) | 64 | 45k | **63.9** | **55.0** | [config](projects/ViTDet/configs/COCO/cascade_mask_rcnn_vitdet_eva.py) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_coco_seg.pth) (`4GB`) |
104
+
105
+ </div>
106
+
107
+ ### LVIS v1.0 (single-scale evaluation on `val` set)
108
+
109
+ <div align="center">
110
+
111
+ | init. model weight | batch size | iter | AP box | AP mask | config | model weight |
112
+ | :---: | :---: |:----:|:--------:|:---------------:|:---------------------------------------:|:--------------------------------------------------------------------:|
113
+ | [`eva_o365`](https://huggingface.co/BAAI/EVA/blob/main/eva_o365.pth) | 64 | 75k | **62.2** | **55.0**| [config](projects/ViTDet/configs/LVIS/cascade_mask_rcnn_vitdet_eva.py) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_lvis.pth) (`4GB`) |
114
+
115
+ </div>
116
+
117
+
118
+ ## Semantic Segmentation
119
+
120
+ ### COCO-Stuff-164K
121
+
122
+ <div align="center">
123
+
124
+ | init. model weight | batch size | iter | crop size | mIoU (ss) | config | seg model weight |logs|
125
+ | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
126
+ | [`eva_psz14to16`](https://huggingface.co/BAAI/EVA/blob/main/eva_psz14to16.pt) | 32 | 60k | 896 | **53.4** | [config](configs/coco_stuff164k/eva_mask2former_896_60k_cocostuff164k_ss.py) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_sem_seg_mask2former_cocostuff_53p4.pth) | [training](../logs/sem_seg/ft_cocstuff164k_sem_seg_ss_53p4_training_log.txt) \| [evaluation](../logs/sem_seg/ft_cocstuff164k_sem_seg_ss_53p4.txt)
127
+
128
+ </div>
129
+
130
+ ### ADE20K
131
+
132
+ <div align="center">
133
+
134
+ | init. model weight | batch size | iter | crop size | mIoU | config | seg model weight |logs|
135
+ | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
136
+ | [`eva_sem_seg_coco`](https://huggingface.co/BAAI/EVA/blob/main/eva_sem_seg_mask2former_cocostuff_53p4.pth) | 64 | 20k | 896 | **61.5** (ss) \| **62.3** (ms) | [config](configs/ade20k/eva_mask2former_896_20k_coco164k2ade20k_ss.py) | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_sem_seg_mask2former_ade_ss61p5_ms62p3.pth) | [training](../logs/sem_seg/ft_ade20k_sem_seg_ms_62p3_training_log.txt) \| [evaluation](../logs/sem_seg/ft_ade20k_sem_seg_ms_62p3.txt)
137
+
138
+ </div>
139
+
140
+
141
+
142
+ ## EVA-CLIP
143
+
144
+
145
+ <div align="center">
146
+
147
+ | model name | #param. | precision | data | batch size | IN-1K zero-shot top-1 | weight |
148
+ |:-----------:|:------:|:------:|:------:|:------:|:------:|:------:|
149
+ | `eva_clip_psz14` | 1.3B | `fp16` | [LAION-400M](https://laion.ai/laion-400-open-dataset/) | 41K | **78.5** | [🤗 HF link](https://huggingface.co/BAAI/EVA/blob/main/eva_clip_psz14.pt) (`2GB`) |
150
+
151
+ </div>
152
+
153
+ > The ImageNet-1K zero-shot classification performance is higher than our paper (`78.5` *v.s.* `78.2`) because of longer training.
154
+
155
+ We choose to train a 1.3B CLIP model, not because it is easy, but because it is hard. Please refer to [this note](https://docs.google.com/document/d/1FXosAZ3wMrzThgnWR6KSkXIz4IMItq3umDGos38pJps/edit) for a glance of the challenges in training very large CLIP.
156
+
157
+ To our knowledge, EVA-CLIP is **the largest performant open-sourced CLIP model** evaluated via zero-shot classification performance.
158
+ We will updates the results in our paper soon.
159
+ For more details of EVA-CLIP, please refer to Section 2.3.5 of [our paper](https://arxiv.org/pdf/2211.07636.pdf).
160
+
161
+ We hope open-sourcing EVA-CLIP can facilitate future research in multi-modal learning, representation leaning, AIGC, *etc*.
162
+
163
+
164
+
165
+ ## Citation
166
+ If you find our work helpful, please star this repo and cite the related articles. Thanks for your support!
167
+
168
+ ```
169
+ @article{EVA,
170
+ title={EVA: Exploring the Limits of Masked Visual Representation Learning at Scale},
171
+ author={Fang, Yuxin and Wang, Wen and Xie, Binhui and Sun, Quan and Wu, Ledell and Wang, Xinggang and Huang, Tiejun and Wang, Xinlong and Cao, Yue},
172
+ journal={arXiv preprint arXiv:2211.07636},
173
+ year={2022}
174
+ }
175
+ ```
176
+
177
+
178
+ ## License
179
+
180
+ The content of this project itself is licensed under the MIT License.
181
+
182
+ ## Contact
183
+
184
+ For help or issues using EVA, please open a GitHub [issue](https://github.com/baaivision/EVA/issues/new).
185
+
186
+ **We are hiring** at all levels at BAAI Vision Team, including full-time researchers, engineers and interns.
187
+ If you are interested in working with us on **foundation model, self-supervised learning and multimodal learning**, please contact [Yue Cao](http://yue-cao.me/) (`caoyue@baai.ac.cn`) and [Xinlong Wang](https://www.xloong.wang/) (`wangxinlong@baai.ac.cn`).
188
+