Update README.md
Browse files
README.md
CHANGED
@@ -1,111 +1,111 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
tags:
|
4 |
-
- vision
|
5 |
-
- DPO
|
6 |
-
- RLHF
|
7 |
-
- preference
|
8 |
-
- feedback
|
9 |
-
- reward model
|
10 |
-
- preference model
|
11 |
-
---
|
12 |
-
|
13 |
-
#### Robust Visual Reward Model
|
14 |
-
Robust visual reward model (RoVRM) is developed through a three-phase progressive training (i.e., pre-training with textual preference data→fine-tuning with image caption-based preference data→fine-tuning with visual preference data), and optimal transport-based selective preference data.
|
15 |
-
These approaches effectively transfer preferences from auxiliary textual data to enhance the model's robustness.
|
16 |
-
The repository hosts the RoVRM built on the LLaVA-1.5-7B model.
|
17 |
-
We employed RoVRM for best-of-$n$ sampling and RL training, demonstrating its capability to significantly improve performance and reduce hallucination in large vision-language models.
|
18 |
-
Detailed training information and experimental results are available in our [paper](https://arxiv.org/abs/2408.12109).
|
19 |
-
|
20 |
-

|
21 |
-
|
22 |
-
## How to use the model
|
23 |
-
We recommend using the [Vision-LLM-Alignment](https://github.com/wangclnlp/Vision-LLM-Alignment) system to run our RoVRM, as it was also used for its training.
|
24 |
-
|
25 |
-
To evaluate a question-answer pair with RoVRM, follow two steps:
|
26 |
-
1. Convert the safetensor format model to ```pytorch_model.bin``` by using the ```convert_pytorch_bin.py``` script.
|
27 |
-
2. Download the Vision-LLM-Alignment repository and run the demo from the first-level directory within the repository.
|
28 |
-
|
29 |
-
```python
|
30 |
-
from transformers import AutoProcessor
|
31 |
-
from training.utils.model.third_party_model.hf_model.modeling_llava import LlavaForConditionalGeneration
|
32 |
-
from torch.utils.data.dataloader import default_collate
|
33 |
-
from PIL import Image
|
34 |
-
import copy
|
35 |
-
import torch
|
36 |
-
import argparse
|
37 |
-
import os
|
38 |
-
|
39 |
-
device = torch.device("cuda:0")
|
40 |
-
from training.utils.model.modeling_reward import create_reward_or_critic_model
|
41 |
-
|
42 |
-
# Set vis_llm_base path and path of the checkpoint
|
43 |
-
# You need to load the llava-1.5-7b model to build an initialized RoVRM.
|
44 |
-
base_path = "base_models/llava-1.5-7b-hf"
|
45 |
-
|
46 |
-
# the checkpoint of RoVRM.
|
47 |
-
ckpt_path = "models/pytorch_model.bin"
|
48 |
-
|
49 |
-
processor = AutoProcessor.from_pretrained(base_path)
|
50 |
-
image_processor = processor.image_processor
|
51 |
-
tokenizer = processor.tokenizer
|
52 |
-
|
53 |
-
tokenizer.add_bos_token = True
|
54 |
-
tokenizer.add_eos_token = True
|
55 |
-
|
56 |
-
args = {
|
57 |
-
"model_architecture": "llava",
|
58 |
-
"lang_decoder_update": False,
|
59 |
-
"from_checkpoint": base_path
|
60 |
-
}
|
61 |
-
args = argparse.Namespace(**args)
|
62 |
-
|
63 |
-
model, image_processor, tokenizer = create_reward_or_critic_model(
|
64 |
-
text_tokenizer=tokenizer,
|
65 |
-
args=args)
|
66 |
-
|
67 |
-
model.load_state_dict(torch.load(os.path.join(ckpt_path, 'pytorch_model.bin'), map_location='cpu'), strict=False)
|
68 |
-
model.to(device)
|
69 |
-
|
70 |
-
# Set input sentence and path of the input image
|
71 |
-
# <image> is necessary when there is an image input
|
72 |
-
input_sen = "USER: ### Image:<image>\nIdentify and describe each object in the image in detail.\nASSISTANT: In the image, there is a cute, colorful cartoon girl sitting on a chair at a wooden table. She is reading a book, which is a prominent object in the scene. The table and chair are also present, adding to the overall setting. As this is a cartoon-style image, the girl and the book may have a more exaggerated or simplified design compared to real-life objects. "
|
73 |
-
img_path = "llava1.5_raw_images_00011_000118793.jpg"
|
74 |
-
|
75 |
-
# Load and preprocess the image
|
76 |
-
image = Image.open(img_path).convert("RGB")
|
77 |
-
image = image_processor(image)
|
78 |
-
try:
|
79 |
-
image = image['pixel_values'][0]
|
80 |
-
except:
|
81 |
-
pass
|
82 |
-
|
83 |
-
input_sen = tokenizer(input_sen,
|
84 |
-
return_tensors=None,
|
85 |
-
padding="do_not_pad",
|
86 |
-
truncation=True,
|
87 |
-
max_length=512,)
|
88 |
-
|
89 |
-
input_sen.update(labels=copy.deepcopy(input_sen["input_ids"]))
|
90 |
-
input_sen.update(image=image)
|
91 |
-
|
92 |
-
reward_scores = model(img=default_collate(image).reshape((-1,) + image[0].shape[-3:]).unsqueeze(0).to(device),
|
93 |
-
lang=torch.LongTensor(input_sen["input_ids"]).unsqueeze(0).to(device),
|
94 |
-
attention_mask=torch.LongTensor(input_sen["attention_mask"]).unsqueeze(0).to(device),
|
95 |
-
input_labels=torch.LongTensor(input_sen["labels"]).unsqueeze(0).to(device))
|
96 |
-
|
97 |
-
print(reward_scores[0].item())
|
98 |
-
```
|
99 |
-
|
100 |
-
Please cite our paper if you find RoVRM helpful in your work🌹🌹🌹:
|
101 |
-
```
|
102 |
-
@
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
}
|
111 |
```
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
tags:
|
4 |
+
- vision
|
5 |
+
- DPO
|
6 |
+
- RLHF
|
7 |
+
- preference
|
8 |
+
- feedback
|
9 |
+
- reward model
|
10 |
+
- preference model
|
11 |
+
---
|
12 |
+
|
13 |
+
#### Robust Visual Reward Model
|
14 |
+
Robust visual reward model (RoVRM) is developed through a three-phase progressive training (i.e., pre-training with textual preference data→fine-tuning with image caption-based preference data→fine-tuning with visual preference data), and optimal transport-based selective preference data.
|
15 |
+
These approaches effectively transfer preferences from auxiliary textual data to enhance the model's robustness.
|
16 |
+
The repository hosts the RoVRM built on the LLaVA-1.5-7B model.
|
17 |
+
We employed RoVRM for best-of-$n$ sampling and RL training, demonstrating its capability to significantly improve performance and reduce hallucination in large vision-language models.
|
18 |
+
Detailed training information and experimental results are available in our [paper](https://arxiv.org/abs/2408.12109).
|
19 |
+
|
20 |
+

|
21 |
+
|
22 |
+
## How to use the model
|
23 |
+
We recommend using the [Vision-LLM-Alignment](https://github.com/wangclnlp/Vision-LLM-Alignment) system to run our RoVRM, as it was also used for its training.
|
24 |
+
|
25 |
+
To evaluate a question-answer pair with RoVRM, follow two steps:
|
26 |
+
1. Convert the safetensor format model to ```pytorch_model.bin``` by using the ```convert_pytorch_bin.py``` script.
|
27 |
+
2. Download the Vision-LLM-Alignment repository and run the demo from the first-level directory within the repository.
|
28 |
+
|
29 |
+
```python
|
30 |
+
from transformers import AutoProcessor
|
31 |
+
from training.utils.model.third_party_model.hf_model.modeling_llava import LlavaForConditionalGeneration
|
32 |
+
from torch.utils.data.dataloader import default_collate
|
33 |
+
from PIL import Image
|
34 |
+
import copy
|
35 |
+
import torch
|
36 |
+
import argparse
|
37 |
+
import os
|
38 |
+
|
39 |
+
device = torch.device("cuda:0")
|
40 |
+
from training.utils.model.modeling_reward import create_reward_or_critic_model
|
41 |
+
|
42 |
+
# Set vis_llm_base path and path of the checkpoint
|
43 |
+
# You need to load the llava-1.5-7b model to build an initialized RoVRM.
|
44 |
+
base_path = "base_models/llava-1.5-7b-hf"
|
45 |
+
|
46 |
+
# the checkpoint of RoVRM.
|
47 |
+
ckpt_path = "models/pytorch_model.bin"
|
48 |
+
|
49 |
+
processor = AutoProcessor.from_pretrained(base_path)
|
50 |
+
image_processor = processor.image_processor
|
51 |
+
tokenizer = processor.tokenizer
|
52 |
+
|
53 |
+
tokenizer.add_bos_token = True
|
54 |
+
tokenizer.add_eos_token = True
|
55 |
+
|
56 |
+
args = {
|
57 |
+
"model_architecture": "llava",
|
58 |
+
"lang_decoder_update": False,
|
59 |
+
"from_checkpoint": base_path
|
60 |
+
}
|
61 |
+
args = argparse.Namespace(**args)
|
62 |
+
|
63 |
+
model, image_processor, tokenizer = create_reward_or_critic_model(
|
64 |
+
text_tokenizer=tokenizer,
|
65 |
+
args=args)
|
66 |
+
|
67 |
+
model.load_state_dict(torch.load(os.path.join(ckpt_path, 'pytorch_model.bin'), map_location='cpu'), strict=False)
|
68 |
+
model.to(device)
|
69 |
+
|
70 |
+
# Set input sentence and path of the input image
|
71 |
+
# <image> is necessary when there is an image input
|
72 |
+
input_sen = "USER: ### Image:<image>\nIdentify and describe each object in the image in detail.\nASSISTANT: In the image, there is a cute, colorful cartoon girl sitting on a chair at a wooden table. She is reading a book, which is a prominent object in the scene. The table and chair are also present, adding to the overall setting. As this is a cartoon-style image, the girl and the book may have a more exaggerated or simplified design compared to real-life objects. "
|
73 |
+
img_path = "llava1.5_raw_images_00011_000118793.jpg"
|
74 |
+
|
75 |
+
# Load and preprocess the image
|
76 |
+
image = Image.open(img_path).convert("RGB")
|
77 |
+
image = image_processor(image)
|
78 |
+
try:
|
79 |
+
image = image['pixel_values'][0]
|
80 |
+
except:
|
81 |
+
pass
|
82 |
+
|
83 |
+
input_sen = tokenizer(input_sen,
|
84 |
+
return_tensors=None,
|
85 |
+
padding="do_not_pad",
|
86 |
+
truncation=True,
|
87 |
+
max_length=512,)
|
88 |
+
|
89 |
+
input_sen.update(labels=copy.deepcopy(input_sen["input_ids"]))
|
90 |
+
input_sen.update(image=image)
|
91 |
+
|
92 |
+
reward_scores = model(img=default_collate(image).reshape((-1,) + image[0].shape[-3:]).unsqueeze(0).to(device),
|
93 |
+
lang=torch.LongTensor(input_sen["input_ids"]).unsqueeze(0).to(device),
|
94 |
+
attention_mask=torch.LongTensor(input_sen["attention_mask"]).unsqueeze(0).to(device),
|
95 |
+
input_labels=torch.LongTensor(input_sen["labels"]).unsqueeze(0).to(device))
|
96 |
+
|
97 |
+
print(reward_scores[0].item())
|
98 |
+
```
|
99 |
+
|
100 |
+
Please cite our paper if you find RoVRM helpful in your work🌹🌹🌹:
|
101 |
+
```
|
102 |
+
@inproceedings{wang2025rovrm,
|
103 |
+
title={Rovrm: A robust visual reward model optimized via auxiliary textual preference data},
|
104 |
+
author={Wang, Chenglong and Gan, Yang and Huo, Yifu and Mu, Yongyu and Yang, Murun and He, Qiaozhi and Xiao, Tong and Zhang, Chunliang and Liu, Tongran and Zhu, Jingbo},
|
105 |
+
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
|
106 |
+
volume={39},
|
107 |
+
number={24},
|
108 |
+
pages={25336--25344},
|
109 |
+
year={2025}
|
110 |
+
}
|
111 |
```
|