许可证: mit
语言:

• 英文
  基础模型:
• Qwen/Qwen2.5-VL-7B-Instruct
  任务标签: 强化学习
  标签:
• 图像质量评估(IQA)
• 视觉语言模型(VLM)
• 推理引导
• Pytorch

最新版本请访问 VisualQuality-R1-7B

VisualQuality-R1-7B预览版

这是VisualQuality-R1的演示版本，基于KADID-10K、TID2013和KONIQ-10K数据集联合训练。其基础模型为Qwen2.5-VL-7B-Instruct。

论文链接: arXiv

快速开始

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor  
from qwen_vl_utils import process_vision_info  

import json  
import numpy as np  
import torch  
import random  
import re  
import os  

def score_image(model_path, image_path):  
    model = Qwen2_5_VLForConditionalGeneration.from_pretrained(  
        model_path,  
        torch_dtype=torch.bfloat16,  
        attn_implementation="flash_attention_2",  
        device_map=device,  
    )  
    processor = AutoProcessor.from_pretrained(MODEL_PATH)  
    processor.tokenizer.padding_side = "left"  

    PROMPT = (  
        "您正在执行图像质量评估任务。问题如下："  
        "您对这张图片的整体质量评分是多少？评分应为1到5之间的浮点数，"  
        "保留两位小数，1代表质量极差，5代表质量极佳。"  
    )  

    x = {  
        "image": [image_path],  
        "question": PROMPT,  
    }  

    QUESTION_TEMPLATE = "{Question} 首先在<think></think>标签中输出思考过程，然后在<answer></answer>标签中仅输出最终评分。"  
    message = [  
        {  
            "role": "用户",  
            "content": [  
                *({'type': 'image', 'image': img_path} for img_path in x['image']),  
                {"type": "text", "text": QUESTION_TEMPLATE.format(Question=x['question'])}  
            ],  
        }  
    ]  

    batch_messages = [message]  

    # 推理准备  
    text = [processor.apply_chat_template(msg, tokenize=False, add_generation_prompt=True, add_vision_id=True) for msg in batch_messages]  
    image_inputs, video_inputs = process_vision_info(batch_messages)  
    inputs = processor(  
        text=text,  
        images=image_inputs,  
        videos=video_inputs,  
        padding=True,  
        return_tensors="pt",  
    )  
    inputs = inputs.to(device)  

    # 推理：生成输出  
    generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=256, do_sample=True)  
    generated_ids_trimmed = [  
        out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)  
    ]  
    batch_output_text = processor.batch_decode(  
        generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False  
    )  

    reasoning = re.findall(r'<think>(.*?)</think>', batch_output_text[0], re.DOTALL)  
    reasoning = reasoning[-1].strip()  

    model_output_matches = re.findall(r'<answer>(.*?)</answer>', batch_output_text[0], re.DOTALL)  
    model_answer = model_output_matches[-1].strip()  
    score = float(re.search(r'\d+(\.\d+)?', model_answer).group())  

    return reasoning, score  

random.seed(42)  
device = torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")  

### 在此修改  
model_path = ""  
image_path = ""  

reasoning, score = score_image(  
    model_path=model_path,  
    image_path=image_path  
)  

print(reasoning)  
print(score)