Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,332 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
base_model:
|
6 |
+
- Qwen/Qwen2.5-VL-7B-Instruct
|
7 |
+
pipeline_tag: reinforcement-learning
|
8 |
+
tags:
|
9 |
+
- IQA
|
10 |
+
- Reasoning
|
11 |
+
- VLM
|
12 |
+
- Pytorch
|
13 |
+
- R1
|
14 |
+
---
|
15 |
+
|
16 |
+
# VisualQuality-R1-7B
|
17 |
+
This is the final version of VisualQuality-R1, trained on a diverse combination of synthetic and realistic datasets.<br>
|
18 |
+
Paper link: [arXiv](https://arxiv.org/abs/2505.14460)<br>
|
19 |
+
Code link: [github](https://github.com/TianheWu/VisualQuality-R1)
|
20 |
+
|
21 |
+
> The first NR-IQA model enhanced by RL2R, capable of both quality description and rating through reasoning.
|
22 |
+
|
23 |
+
|
24 |
+
|
25 |
+

|
26 |
+
|
27 |
+
|
28 |
+
|
29 |
+
## Quick Start
|
30 |
+
This section includes the usages of **VisualQuality-R1**.
|
31 |
+
|
32 |
+
<details>
|
33 |
+
<summary>Example Code (Single Image Quality Rating)</summary>
|
34 |
+
|
35 |
+
```python
|
36 |
+
from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
|
37 |
+
from qwen_vl_utils import process_vision_info
|
38 |
+
|
39 |
+
import torch
|
40 |
+
import random
|
41 |
+
import re
|
42 |
+
import os
|
43 |
+
|
44 |
+
|
45 |
+
def score_image(image_path, model, processor):
|
46 |
+
PROMPT = (
|
47 |
+
"You are doing the image quality assessment task. Here is the question: "
|
48 |
+
"What is your overall rating on the quality of this picture? The rating should be a float between 1 and 5, "
|
49 |
+
"rounded to two decimal places, with 1 representing very poor quality and 5 representing excellent quality. "
|
50 |
+
"First output the thinking process in <think> </think> tags and then output the final answer with only one score in <answer> </answer> tags."
|
51 |
+
)
|
52 |
+
|
53 |
+
QUESTION_TEMPLATE = "{Question} First output the thinking process in <think> </think> tags and then output the final answer with only one score in <answer> </answer> tags."
|
54 |
+
# QUESTION_TEMPLATE = "Please describe the quality of this image."
|
55 |
+
message = [
|
56 |
+
{
|
57 |
+
"role": "user",
|
58 |
+
"content": [
|
59 |
+
{'type': 'image', 'image': image_path},
|
60 |
+
{"type": "text", "text": PROMPT}
|
61 |
+
],
|
62 |
+
}
|
63 |
+
]
|
64 |
+
|
65 |
+
batch_messages = [message]
|
66 |
+
|
67 |
+
# Preparation for inference
|
68 |
+
text = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True, add_vision_id=True) for msg in batch_messages]
|
69 |
+
image_inputs, video_inputs = process_vision_info(batch_messages)
|
70 |
+
inputs = processor(
|
71 |
+
text=text,
|
72 |
+
images=image_inputs,
|
73 |
+
videos=video_inputs,
|
74 |
+
padding=True,
|
75 |
+
return_tensors="pt",
|
76 |
+
)
|
77 |
+
inputs = inputs.to(device)
|
78 |
+
|
79 |
+
# Inference: Generation of the output
|
80 |
+
generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=2048, do_sample=True, top_k=50, top_p=1)
|
81 |
+
generated_ids_trimmed = [
|
82 |
+
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
83 |
+
]
|
84 |
+
batch_output_text = processor.batch_decode(
|
85 |
+
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
86 |
+
)
|
87 |
+
|
88 |
+
reasoning = re.findall(r'<think>(.*?)</think>', batch_output_text[0], re.DOTALL)
|
89 |
+
reasoning = reasoning[-1].strip()
|
90 |
+
|
91 |
+
try:
|
92 |
+
model_output_matches = re.findall(r'<answer>(.*?)</answer>', batch_output_text[0], re.DOTALL)
|
93 |
+
model_answer = model_output_matches[-1].strip() if model_output_matches else batch_output_text[0].strip()
|
94 |
+
score = float(re.search(r'\d+(\.\d+)?', model_answer).group())
|
95 |
+
except:
|
96 |
+
print(f"================= Meet error with {img_path}, please generate again. =================")
|
97 |
+
score = random.randint(1, 5)
|
98 |
+
|
99 |
+
return reasoning, score
|
100 |
+
|
101 |
+
|
102 |
+
random.seed(1)
|
103 |
+
MODEL_PATH = ""
|
104 |
+
device = torch.device("cuda:5") if torch.cuda.is_available() else torch.device("cpu")
|
105 |
+
image_path = ""
|
106 |
+
|
107 |
+
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
108 |
+
MODEL_PATH,
|
109 |
+
torch_dtype=torch.bfloat16,
|
110 |
+
attn_implementation="flash_attention_2",
|
111 |
+
device_map=device,
|
112 |
+
)
|
113 |
+
processor = AutoProcessor.from_pretrained(MODEL_PATH)
|
114 |
+
processor.tokenizer.padding_side = "left"
|
115 |
+
|
116 |
+
reasoning, score = score_image(
|
117 |
+
image_path, model, processor
|
118 |
+
)
|
119 |
+
|
120 |
+
print(reasoning)
|
121 |
+
print(score)
|
122 |
+
```
|
123 |
+
</details>
|
124 |
+
|
125 |
+
|
126 |
+
<details>
|
127 |
+
<summary>Example Code (Batch Images Quality Rating)</summary>
|
128 |
+
|
129 |
+
```python
|
130 |
+
from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
|
131 |
+
from qwen_vl_utils import process_vision_info
|
132 |
+
from tqdm import tqdm
|
133 |
+
|
134 |
+
import torch
|
135 |
+
import random
|
136 |
+
import re
|
137 |
+
import os
|
138 |
+
|
139 |
+
|
140 |
+
def get_image_paths(folder_path):
|
141 |
+
image_extensions = {'.jpg', '.jpeg', '.png', '.bmp', '.gif', '.tiff', '.webp'}
|
142 |
+
image_paths = []
|
143 |
+
|
144 |
+
for root, dirs, files in os.walk(folder_path):
|
145 |
+
for file in files:
|
146 |
+
_, ext = os.path.splitext(file)
|
147 |
+
if ext.lower() in image_extensions:
|
148 |
+
image_paths.append(os.path.join(root, file))
|
149 |
+
|
150 |
+
return image_paths
|
151 |
+
|
152 |
+
def score_batch_image(image_paths, model, processor):
|
153 |
+
PROMPT = (
|
154 |
+
"You are doing the image quality assessment task. Here is the question: "
|
155 |
+
"What is your overall rating on the quality of this picture? The rating should be a float between 1 and 5, "
|
156 |
+
"rounded to two decimal places, with 1 representing very poor quality and 5 representing excellent quality."
|
157 |
+
)
|
158 |
+
|
159 |
+
QUESTION_TEMPLATE = "{Question} First output the thinking process in <think> </think> tags and then output the final answer with only one score in <answer> </answer> tags."
|
160 |
+
|
161 |
+
messages = []
|
162 |
+
for img_path in image_paths:
|
163 |
+
message = [
|
164 |
+
{
|
165 |
+
"role": "user",
|
166 |
+
"content": [
|
167 |
+
{'type': 'image', 'image': img_path},
|
168 |
+
{"type": "text", "text": QUESTION_TEMPLATE.format(Question=PROMPT)}
|
169 |
+
],
|
170 |
+
}
|
171 |
+
]
|
172 |
+
messages.append(message)
|
173 |
+
|
174 |
+
BSZ = 32
|
175 |
+
all_outputs = [] # List to store all answers
|
176 |
+
for i in tqdm(range(0, len(messages), BSZ)):
|
177 |
+
batch_messages = messages[i:i + BSZ]
|
178 |
+
|
179 |
+
# Preparation for inference
|
180 |
+
text = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True, add_vision_id=True) for msg in batch_messages]
|
181 |
+
|
182 |
+
image_inputs, video_inputs = process_vision_info(batch_messages)
|
183 |
+
inputs = processor(
|
184 |
+
text=text,
|
185 |
+
images=image_inputs,
|
186 |
+
videos=video_inputs,
|
187 |
+
padding=True,
|
188 |
+
return_tensors="pt",
|
189 |
+
)
|
190 |
+
inputs = inputs.to(device)
|
191 |
+
|
192 |
+
# Inference: Generation of the output
|
193 |
+
generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=512, do_sample=True, top_k=50, top_p=1)
|
194 |
+
generated_ids_trimmed = [
|
195 |
+
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
196 |
+
]
|
197 |
+
batch_output_text = processor.batch_decode(
|
198 |
+
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
199 |
+
)
|
200 |
+
|
201 |
+
all_outputs.extend(batch_output_text)
|
202 |
+
|
203 |
+
path_score_dict = {}
|
204 |
+
for img_path, model_output in zip(image_paths, all_outputs):
|
205 |
+
reasoning = re.findall(r'<think>(.*?)</think>', model_output, re.DOTALL)
|
206 |
+
reasoning = reasoning[-1].strip()
|
207 |
+
|
208 |
+
try:
|
209 |
+
model_output_matches = re.findall(r'<answer>(.*?)</answer>', model_output, re.DOTALL)
|
210 |
+
model_answer = model_output_matches[-1].strip() if model_output_matches else model_output.strip()
|
211 |
+
score = float(re.search(r'\d+(\.\d+)?', model_answer).group())
|
212 |
+
except:
|
213 |
+
print(f"Meet error with {img_path}, please generate again.")
|
214 |
+
score = random.randint(1, 5)
|
215 |
+
|
216 |
+
path_score_dict[img_path] = score
|
217 |
+
|
218 |
+
return path_score_dict
|
219 |
+
|
220 |
+
|
221 |
+
random.seed(1)
|
222 |
+
MODEL_PATH = ""
|
223 |
+
device = torch.device("cuda:3") if torch.cuda.is_available() else torch.device("cpu")
|
224 |
+
|
225 |
+
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
226 |
+
MODEL_PATH,
|
227 |
+
torch_dtype=torch.bfloat16,
|
228 |
+
attn_implementation="flash_attention_2",
|
229 |
+
device_map=device,
|
230 |
+
)
|
231 |
+
processor = AutoProcessor.from_pretrained(MODEL_PATH)
|
232 |
+
processor.tokenizer.padding_side = "left"
|
233 |
+
|
234 |
+
image_root = ""
|
235 |
+
image_paths = get_image_paths(image_root) # It should be a list
|
236 |
+
|
237 |
+
path_score_dict = score_batch_image(
|
238 |
+
image_paths, model, processor
|
239 |
+
)
|
240 |
+
|
241 |
+
file_name = "output.txt"
|
242 |
+
with open(file_name, "w") as file:
|
243 |
+
for key, value in path_score_dict.items():
|
244 |
+
file.write(f"{key} {value}\n")
|
245 |
+
|
246 |
+
print("Done!")
|
247 |
+
```
|
248 |
+
</details>
|
249 |
+
|
250 |
+
|
251 |
+
<details>
|
252 |
+
<summary>Example Code (Images Inference)</summary>
|
253 |
+
|
254 |
+
You can prompt anything what you like in the following commands (including multi-image as input)
|
255 |
+
```python
|
256 |
+
from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
|
257 |
+
from qwen_vl_utils import process_vision_info
|
258 |
+
|
259 |
+
import torch
|
260 |
+
import random
|
261 |
+
import re
|
262 |
+
import os
|
263 |
+
|
264 |
+
|
265 |
+
def generate(image_paths, model, prompt, processor):
|
266 |
+
message = [
|
267 |
+
{
|
268 |
+
"role": "user",
|
269 |
+
"content": [
|
270 |
+
*({'type': 'image', 'image': img_path} for img_path in image_paths),
|
271 |
+
{"type": "text", "text": prompt}
|
272 |
+
],
|
273 |
+
}
|
274 |
+
]
|
275 |
+
|
276 |
+
batch_messages = [message]
|
277 |
+
|
278 |
+
# Preparation for inference
|
279 |
+
text = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True, add_vision_id=True) for msg in batch_messages]
|
280 |
+
image_inputs, video_inputs = process_vision_info(batch_messages)
|
281 |
+
inputs = processor(
|
282 |
+
text=text,
|
283 |
+
images=image_inputs,
|
284 |
+
videos=video_inputs,
|
285 |
+
padding=True,
|
286 |
+
return_tensors="pt",
|
287 |
+
)
|
288 |
+
inputs = inputs.to(device)
|
289 |
+
|
290 |
+
# Inference: Generation of the output
|
291 |
+
generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=2048, do_sample=True, top_k=50, top_p=1)
|
292 |
+
generated_ids_trimmed = [
|
293 |
+
out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
|
294 |
+
]
|
295 |
+
batch_output_text = processor.batch_decode(
|
296 |
+
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
297 |
+
)
|
298 |
+
|
299 |
+
return batch_output_text[0]
|
300 |
+
|
301 |
+
|
302 |
+
random.seed(1)
|
303 |
+
MODEL_PATH = ""
|
304 |
+
device = torch.device("cuda:5") if torch.cuda.is_available() else torch.device("cpu")
|
305 |
+
image_path = [
|
306 |
+
"",
|
307 |
+
""
|
308 |
+
]
|
309 |
+
|
310 |
+
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
311 |
+
MODEL_PATH,
|
312 |
+
torch_dtype=torch.bfloat16,
|
313 |
+
attn_implementation="flash_attention_2",
|
314 |
+
device_map=device,
|
315 |
+
)
|
316 |
+
processor = AutoProcessor.from_pretrained(MODEL_PATH)
|
317 |
+
processor.tokenizer.padding_side = "left"
|
318 |
+
|
319 |
+
prompt = "Please describe the quality of given two images."
|
320 |
+
answer = generate(
|
321 |
+
image_path, model, prompt, processor
|
322 |
+
)
|
323 |
+
|
324 |
+
print(answer)
|
325 |
+
```
|
326 |
+
</details>
|
327 |
+
|
328 |
+
|
329 |
+
|
330 |
+
|
331 |
+
|
332 |
+
|