Update README.md
Browse files
README.md
CHANGED
@@ -3,31 +3,57 @@ For more details please refer to our github repo: https://github.com/FlagOpen/Fl
|
|
3 |
# [Visualized BGE](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/visual)
|
4 |
|
5 |
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
|
8 |
- Multi-Modal Knowledge Retrieval (query: text; candidate: image-text pairs, text, or image) e.g. [WebQA](https://github.com/WebQnA/WebQA)
|
9 |
-
- Composed Image Retrieval (query: image-text pair; candidate: images) e.g. [CIRR](), [FashionIQ]()
|
10 |
-
- Knowledge Retrieval with Multi-Modal Queries (query: image-text pair; candidate: texts) e.g. [ReMuQ]()
|
11 |
|
12 |
Moreover, Visualized BGE fully preserves the strong text embedding capabilities of the original BGE model : )
|
13 |
|
14 |
## Specs
|
15 |
-
|
16 |
### Model
|
17 |
| **Model Name** | **Dimension** | **Text Embedding Model** | **Language** | **Weight** |
|
18 |
| --- | --- | --- | --- | --- |
|
19 |
| BAAI/bge-visualized-base-en-v1.5 | 768 | [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5) | English | [π€ HF link](https://huggingface.co/BAAI/bge-visualized/blob/main/Visualized_base_en_v1.5.pth) |
|
20 |
| BAAI/bge-visualized-m3 | 1024 | [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3) | Multilingual | [π€ HF link](https://huggingface.co/BAAI/bge-visualized/blob/main/Visualized_m3.pth) |
|
21 |
|
|
|
22 |
### Data
|
23 |
-
We have generated a hybrid multi-modal dataset consisting of over 500,000 instances for training.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
## Usage
|
26 |
### Installation:
|
27 |
#### Install FlagEmbedding:
|
28 |
```
|
29 |
git clone https://github.com/FlagOpen/FlagEmbedding.git
|
30 |
-
cd FlagEmbedding
|
31 |
pip install -e .
|
32 |
```
|
33 |
#### Another Core Packages:
|
@@ -37,15 +63,15 @@ pip install torchvision timm einops ftfy
|
|
37 |
You don't need to install `xformer` and `apex`. They are not essential for inference and can often cause issues.
|
38 |
|
39 |
### Generate Embedding for Multi-Modal Data:
|
40 |
-
|
41 |
|
42 |
> **Note:** Please download the model weight file ([bge-visualized-base-en-v1.5](https://huggingface.co/BAAI/bge-visualized/resolve/main/Visualized_base_en_v1.5.pth?download=true), [bge-visualized-m3](https://huggingface.co/BAAI/bge-visualized/resolve/main/Visualized_m3.pth?download=true)) in advance and pass the path to the `model_weight` parameter.
|
43 |
|
44 |
-
- Composed Image
|
45 |
``` python
|
46 |
-
|
47 |
import torch
|
48 |
-
from
|
49 |
|
50 |
model = Visualized_BGE(model_name_bge = "BAAI/bge-base-en-v1.5", model_weight="path: Visualized_base_en_v1.5.pth")
|
51 |
model.eval()
|
@@ -63,10 +89,10 @@ print(sim_1, sim_2) # tensor([[0.8750]]) tensor([[0.7816]])
|
|
63 |
``` python
|
64 |
####### Use Visualized BGE doing multi-modal knowledge retrieval
|
65 |
import torch
|
66 |
-
from
|
67 |
|
68 |
model = Visualized_BGE(model_name_bge = "BAAI/bge-base-en-v1.5", model_weight="path: Visualized_base_en_v1.5.pth")
|
69 |
-
|
70 |
with torch.no_grad():
|
71 |
query_emb = model.encode(text="Are there sidewalks on both sides of the Mid-Hudson Bridge?")
|
72 |
candi_emb_1 = model.encode(text="The Mid-Hudson Bridge, spanning the Hudson River between Poughkeepsie and Highland.", image="./imgs/wiki_candi_1.jpg")
|
@@ -78,13 +104,11 @@ sim_2 = query_emb @ candi_emb_2.T
|
|
78 |
sim_3 = query_emb @ candi_emb_3.T
|
79 |
print(sim_1, sim_2, sim_3) # tensor([[0.6932]]) tensor([[0.4441]]) tensor([[0.6415]])
|
80 |
```
|
81 |
-
|
82 |
- Multilingual Multi-Modal Retrieval
|
83 |
``` python
|
84 |
##### Use M3 doing Multilingual Multi-Modal Retrieval
|
85 |
-
|
86 |
import torch
|
87 |
-
from
|
88 |
|
89 |
model = Visualized_BGE(model_name_bge = "BAAI/bge-m3", model_weight="path: Visualized_m3.pth")
|
90 |
model.eval()
|
@@ -97,6 +121,8 @@ sim_1 = query_emb @ candi_emb_1.T
|
|
97 |
sim_2 = query_emb @ candi_emb_2.T
|
98 |
print(sim_1, sim_2) # tensor([[0.7026]]) tensor([[0.8075]])
|
99 |
```
|
|
|
|
|
100 |
|
101 |
## Evaluation Result
|
102 |
Visualized BGE delivers outstanding zero-shot performance across multiple hybrid modal retrieval tasks. It can also serve as a base model for downstream fine-tuning for hybrid modal retrieval tasks.
|
@@ -114,6 +140,9 @@ Visualized BGE delivers outstanding zero-shot performance across multiple hybrid
|
|
114 |
![image.png](./imgs/SFT-CIRR.png)
|
115 |
- Supervised fine-tuning performance on the ReMuQ test set.
|
116 |
![image.png](./imgs/SFT-ReMuQ.png)
|
|
|
|
|
|
|
117 |
## FAQ
|
118 |
|
119 |
**Q1: Can Visualized BGE be used for cross-modal retrieval (text to image)?**
|
@@ -124,6 +153,12 @@ A1: While it is technically possible, it's not the recommended use case. Our mod
|
|
124 |
The image token embedding model in this project is built upon the foundations laid by [EVA-CLIP](https://github.com/baaivision/EVA/tree/master/EVA-CLIP).
|
125 |
|
126 |
## Citation
|
127 |
-
If you find this repository useful, please consider giving a
|
128 |
-
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
# [Visualized BGE](https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/visual)
|
4 |
|
5 |
|
6 |
+
|
7 |
+
## π News
|
8 |
+
**[2024.8.27] The core code for the evaluation and fine-tuning of VISTA can be obtained from [this link](https://github.com/JUNJIE99/VISTA_Evaluation_FineTuning). This includes Stage2 training, downstream task fine-tuning, as well as the datasets we used for evaluation.**
|
9 |
+
|
10 |
+
**[2024.6.13] We have released [VISTA-S2 dataset](https://huggingface.co/datasets/JUNJIE99/VISTA_S2), a hybrid multi-modal dataset consisting of over 500,000 instances for multi-modal training (Stage-2 training in our paper).**
|
11 |
+
|
12 |
+
**[2024.6.7] We have released our paper. [Arxiv Link](https://arxiv.org/abs/2406.04292)**
|
13 |
+
|
14 |
+
**[2024.3.18] We have released our code and model.**
|
15 |
+
|
16 |
+
|
17 |
+
|
18 |
+
|
19 |
+
## Introduction
|
20 |
+
In this project, we introduce Visualized-BGE, a universal multi-modal embedding model. By incorporating image token embedding into the BGE Text Embedding framework, Visualized-BGE gains the flexibility to process multi-modal data that goes beyond just text. Visualized-BGE is mainly used for hybrid modal retrieval tasks, including but not limited to:
|
21 |
|
22 |
- Multi-Modal Knowledge Retrieval (query: text; candidate: image-text pairs, text, or image) e.g. [WebQA](https://github.com/WebQnA/WebQA)
|
23 |
+
- Composed Image Retrieval (query: image-text pair; candidate: images) e.g. [CIRR](https://github.com/Cuberick-Orion/CIRR), [FashionIQ](https://github.com/XiaoxiaoGuo/fashion-iq)
|
24 |
+
- Knowledge Retrieval with Multi-Modal Queries (query: image-text pair; candidate: texts) e.g. [ReMuQ](https://github.com/luomancs/ReMuQ)
|
25 |
|
26 |
Moreover, Visualized BGE fully preserves the strong text embedding capabilities of the original BGE model : )
|
27 |
|
28 |
## Specs
|
|
|
29 |
### Model
|
30 |
| **Model Name** | **Dimension** | **Text Embedding Model** | **Language** | **Weight** |
|
31 |
| --- | --- | --- | --- | --- |
|
32 |
| BAAI/bge-visualized-base-en-v1.5 | 768 | [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5) | English | [π€ HF link](https://huggingface.co/BAAI/bge-visualized/blob/main/Visualized_base_en_v1.5.pth) |
|
33 |
| BAAI/bge-visualized-m3 | 1024 | [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3) | Multilingual | [π€ HF link](https://huggingface.co/BAAI/bge-visualized/blob/main/Visualized_m3.pth) |
|
34 |
|
35 |
+
|
36 |
### Data
|
37 |
+
We have generated a hybrid multi-modal dataset consisting of over 500,000 instances for multi-modal training (Stage-2 training in our paper). You can download our dataset from this [π€ HF Link](https://huggingface.co/datasets/JUNJIE99/VISTA_S2).
|
38 |
+
Process the image compression package with the following commands:
|
39 |
+
|
40 |
+
```bash
|
41 |
+
cat images.tar.part* > images.tar
|
42 |
+
tar -xvf images.tar
|
43 |
+
```
|
44 |
+
If you obtain the following directory structure. You can then use the annotation information (json files) for your own training:
|
45 |
+
```
|
46 |
+
images
|
47 |
+
|__coco
|
48 |
+
|__edit_image
|
49 |
+
```
|
50 |
|
51 |
## Usage
|
52 |
### Installation:
|
53 |
#### Install FlagEmbedding:
|
54 |
```
|
55 |
git clone https://github.com/FlagOpen/FlagEmbedding.git
|
56 |
+
cd FlagEmbedding/research/visual_bge
|
57 |
pip install -e .
|
58 |
```
|
59 |
#### Another Core Packages:
|
|
|
63 |
You don't need to install `xformer` and `apex`. They are not essential for inference and can often cause issues.
|
64 |
|
65 |
### Generate Embedding for Multi-Modal Data:
|
66 |
+
Visualized-BGE provides the versatility to encode multi-modal data in a variety of formats, whether it's purely text, solely image-based, or a combination of both.
|
67 |
|
68 |
> **Note:** Please download the model weight file ([bge-visualized-base-en-v1.5](https://huggingface.co/BAAI/bge-visualized/resolve/main/Visualized_base_en_v1.5.pth?download=true), [bge-visualized-m3](https://huggingface.co/BAAI/bge-visualized/resolve/main/Visualized_m3.pth?download=true)) in advance and pass the path to the `model_weight` parameter.
|
69 |
|
70 |
+
- Composed Image Retrieval
|
71 |
``` python
|
72 |
+
####### Use Visualized BGE doing composed image retrieval
|
73 |
import torch
|
74 |
+
from visual_bge.modeling import Visualized_BGE
|
75 |
|
76 |
model = Visualized_BGE(model_name_bge = "BAAI/bge-base-en-v1.5", model_weight="path: Visualized_base_en_v1.5.pth")
|
77 |
model.eval()
|
|
|
89 |
``` python
|
90 |
####### Use Visualized BGE doing multi-modal knowledge retrieval
|
91 |
import torch
|
92 |
+
from visual_bge.modeling import Visualized_BGE
|
93 |
|
94 |
model = Visualized_BGE(model_name_bge = "BAAI/bge-base-en-v1.5", model_weight="path: Visualized_base_en_v1.5.pth")
|
95 |
+
model.eval()
|
96 |
with torch.no_grad():
|
97 |
query_emb = model.encode(text="Are there sidewalks on both sides of the Mid-Hudson Bridge?")
|
98 |
candi_emb_1 = model.encode(text="The Mid-Hudson Bridge, spanning the Hudson River between Poughkeepsie and Highland.", image="./imgs/wiki_candi_1.jpg")
|
|
|
104 |
sim_3 = query_emb @ candi_emb_3.T
|
105 |
print(sim_1, sim_2, sim_3) # tensor([[0.6932]]) tensor([[0.4441]]) tensor([[0.6415]])
|
106 |
```
|
|
|
107 |
- Multilingual Multi-Modal Retrieval
|
108 |
``` python
|
109 |
##### Use M3 doing Multilingual Multi-Modal Retrieval
|
|
|
110 |
import torch
|
111 |
+
from visual_bge.modeling import Visualized_BGE
|
112 |
|
113 |
model = Visualized_BGE(model_name_bge = "BAAI/bge-m3", model_weight="path: Visualized_m3.pth")
|
114 |
model.eval()
|
|
|
121 |
sim_2 = query_emb @ candi_emb_2.T
|
122 |
print(sim_1, sim_2) # tensor([[0.7026]]) tensor([[0.8075]])
|
123 |
```
|
124 |
+
## Downstream Application Cases
|
125 |
+
- [Huixiangdou](https://github.com/InternLM/HuixiangDou): Using Visualized BGE for the group chat assistant.
|
126 |
|
127 |
## Evaluation Result
|
128 |
Visualized BGE delivers outstanding zero-shot performance across multiple hybrid modal retrieval tasks. It can also serve as a base model for downstream fine-tuning for hybrid modal retrieval tasks.
|
|
|
140 |
![image.png](./imgs/SFT-CIRR.png)
|
141 |
- Supervised fine-tuning performance on the ReMuQ test set.
|
142 |
![image.png](./imgs/SFT-ReMuQ.png)
|
143 |
+
|
144 |
+
|
145 |
+
|
146 |
## FAQ
|
147 |
|
148 |
**Q1: Can Visualized BGE be used for cross-modal retrieval (text to image)?**
|
|
|
153 |
The image token embedding model in this project is built upon the foundations laid by [EVA-CLIP](https://github.com/baaivision/EVA/tree/master/EVA-CLIP).
|
154 |
|
155 |
## Citation
|
156 |
+
If you find this repository useful, please consider giving a star β and citation
|
157 |
+
```
|
158 |
+
@article{zhou2024vista,
|
159 |
+
title={VISTA: Visualized Text Embedding For Universal Multi-Modal Retrieval},
|
160 |
+
author={Zhou, Junjie and Liu, Zheng and Xiao, Shitao and Zhao, Bo and Xiong, Yongping},
|
161 |
+
journal={arXiv preprint arXiv:2406.04292},
|
162 |
+
year={2024}
|
163 |
+
}
|
164 |
+
```
|