File size: 5,929 Bytes
31cb198
 
 
0e7276c
56b40e2
31cb198
 
 
 
 
 
 
56b40e2
31cb198
 
14fe9d6
31cb198
bc37cdc
14fe9d6
31cb198
14fe9d6
31cb198
14fe9d6
31cb198
14fe9d6
31cb198
14fe9d6
 
 
 
98ccbcd
14fe9d6
 
 
ab347b1
14fe9d6
ab347b1
14fe9d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
31cb198
 
 
 
 
 
 
59cea48
31cb198
 
59cea48
dfd5e6c
 
 
59cea48
31cb198
59cea48
31cb198
59cea48
 
 
31cb198
 
59cea48
31cb198
 
 
 
 
 
 
 
 
 
59cea48
31cb198
 
14fe9d6
31cb198
14fe9d6
31cb198
14fe9d6
31cb198
14fe9d6
 
cbd0e39
14fe9d6
 
 
 
 
 
31cb198
14fe9d6
31cb198
14fe9d6
31cb198
14fe9d6
31cb198
 
 
14fe9d6
31cb198
 
14fe9d6
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
---
license: apache-2.0
# inference: false
# pipeline_tag: zero-shot-image-classification
pipeline_tag: feature-extraction

# inference:
#   parameters:
tags:
- clip
- zh
- image-text
- feature-extraction
---

# Taiyi-CLIP-Roberta-102M-Chinese

- Main Page:[Fengshenbang](https://fengshenbang-lm.com/)
- Github: [Fengshenbang-LM](https://github.com/IDEA-CCNL/Fengshenbang-LM)

## 简介 Brief Introduction

首个开源的中文CLIP模型,1.23亿图文对上进行预训练的文本端RoBERTa-base。

The first open source Chinese CLIP, pre-training on 123M image-text pairs, the text encoder: RoBERTa-base.

## 模型分类 Model Taxonomy

|  需求 Demand  | 任务 Task       | 系列 Series      | 模型 Model    | 参数 Parameter | 额外 Extra |
|  :----:  | :----:  | :----:  | :----:  | :----:  | :----:  |
| 特殊 Special | 多模态 Multimodal | 太乙 Taiyi | CLIP (Roberta) |     102M      |    中文 Chinese     |

## 模型信息 Model Information

我们遵循CLIP的实验设置,以获得强大的视觉-语言表征。在训练中文版的CLIP时,我们使用[chinese-roberta-wwm](https://huggingface.co/hfl/chinese-roberta-wwm-ext)作为语言的编码器,并将[CLIP](https://github.com/openai/CLIP)中的ViT-B-32应用于视觉的编码器。为了快速且稳定地进行预训练,我们冻结了视觉编码器并且只微调语言编码器。此外,我们将[Noah-Wukong](https://wukong-dataset.github.io/wukong-dataset/)数据集(100M)和[Zero](https://zero.so.com/)数据集(23M)用作预训练的数据集,训练了24个epoch,在在A100x32上训练了7天。据我们所知,我们的Taiyi-CLIP是目前Huggingface社区中首个的开源中文CLIP。

We follow the experimental setup of CLIP to obtain powerful visual-language intelligence. To obtain the CLIP for Chinese, we employ [chinese-roberta-wwm](https://huggingface.co/hfl/chinese-roberta-wwm-ext) for the language encoder, and apply the ViT-B-32 in [CLIP](https://github.com/openai/CLIP) for the vision encoder. We freeze the vision encoder and tune the language encoder to speed up and stabilize the pre-training process. Moreover, we apply [Noah-Wukong](https://wukong-dataset.github.io/wukong-dataset/) dataset (100M) and [Zero](https://zero.so.com/) dataset (23M) as the pre-training datasets. We train 24 epochs, which takes 7 days to train on A100x16. To the best of our knowledge, our TaiyiCLIP is currently the only open-sourced Chinese CLIP in the huggingface community.

### 下游效果 Performance

**Zero-Shot Classification**

|  model   | dataset  | Top1 | Top5 |
|  ----  | ----  | ---- | ---- |
| Taiyi-CLIP-Roberta-102M-Chinese  | ImageNet1k-CN | 42.85% | 71.48% |

**Zero-Shot Text-to-Image Retrieval**

|  model   | dataset  | Top1 | Top5 | Top10 |
|  ----  | ----  | ---- | ---- | ---- |
| Taiyi-CLIP-Roberta-102M-Chinese  | Flickr30k-CNA-test | 46.32% | 74.58%  | 83.44% |
| Taiyi-CLIP-Roberta-102M-Chinese  | COCO-CN-test | 47.10% | 78.53%  | 87.84% |
| Taiyi-CLIP-Roberta-102M-Chinese  | wukong50k | 49.18% | 81.94% | 90.27% |

## 使用 Usage

```python3
from PIL import Image
import requests
import clip
import torch
from transformers import BertForSequenceClassification, BertConfig, BertTokenizer
from transformers import CLIPProcessor, CLIPModel
import numpy as np

query_texts = ["一只猫", "一只狗",'两只猫', '两只老虎','一只老虎']  # 这里是输入文本的,可以随意替换。
# 加载Taiyi 中文 text encoder
text_tokenizer = BertTokenizer.from_pretrained("IDEA-CCNL/Taiyi-CLIP-Roberta-102M-Chinese")
text_encoder = BertForSequenceClassification.from_pretrained("IDEA-CCNL/Taiyi-CLIP-Roberta-102M-Chinese").eval()
text = text_tokenizer(query_texts, return_tensors='pt', padding=True)['input_ids']

url = "http://images.cocodataset.org/val2017/000000039769.jpg"  # 这里可以换成任意图片的url
# 加载CLIP的image encoder
clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")  
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
image = processor(images=Image.open(requests.get(url, stream=True).raw), return_tensors="pt")

with torch.no_grad():
    image_features = clip_model.get_image_features(**image)
    text_features = text_encoder(text).logits
    # 归一化
    image_features = image_features / image_features.norm(dim=1, keepdim=True)
    text_features = text_features / text_features.norm(dim=1, keepdim=True)
    # 计算余弦相似度 logit_scale是尺度系数
    logit_scale = clip_model.logit_scale.exp()
    logits_per_image = logit_scale * image_features @ text_features.t()
    logits_per_text = logits_per_image.t()
    probs = logits_per_image.softmax(dim=-1).cpu().numpy()
    print(np.around(probs, 3))

```

## 引用 Citation

如果您在您的工作中使用了我们的模型,可以引用我们的[论文](https://arxiv.org/abs/2209.02970):

If you are using the resource for your work, please cite the our [paper](https://arxiv.org/abs/2209.02970):

```text
@article{fengshenbang,
  author    = {Jiaxing Zhang and Ruyi Gan and Junjie Wang and Yuxiang Zhang and Lin Zhang and Ping Yang and Xinyu Gao and Ziwei Wu and Xiaoqun Dong and Junqing He and Jianheng Zhuo and Qi Yang and Yongfeng Huang and Xiayu Li and Yanghan Wu and Junyu Lu and Xinyu Zhu and Weifeng Chen and Ting Han and Kunhao Pan and Rui Wang and Hao Wang and Xiaojun Wu and Zhongshen Zeng and Chongpei Chen},
  title     = {Fengshenbang 1.0: Being the Foundation of Chinese Cognitive Intelligence},
  journal   = {CoRR},
  volume    = {abs/2209.02970},
  year      = {2022}
}
```

也可以引用我们的[网站](https://github.com/IDEA-CCNL/Fengshenbang-LM/):

You can also cite our [website](https://github.com/IDEA-CCNL/Fengshenbang-LM/):

```text
@misc{Fengshenbang-LM,
  title={Fengshenbang-LM},
  author={IDEA-CCNL},
  year={2021},
  howpublished={\url{https://github.com/IDEA-CCNL/Fengshenbang-LM}},
}
```