inference_k2.py

import os
import json
import re
from pathlib import Path
from vllm import LLM, SamplingParams
from transformers import AutoProcessor
from PIL import Image

# --- 1. 环境与硬件配置 ---
os.environ["VLLM_USE_V1"] = "0"
os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
os.environ["PYTHONNOUSERSITE"] = "1"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"


def extract_bboxes_from_response(response_text):
    """从模型回复中提取 bbox 坐标，兼容两种格式：
    GT格式: {"bboxes": ["<x1><y1><x2><y2>", ...]}
    模型输出: [{"bbox": ["<bbox_start><x1><y1><x2><y2><bbox_end>", ...]}]
    """
    try:
        response_clean = response_text.strip()
        if response_clean.startswith('"') and response_clean.endswith('"'):
            try:
                response_clean = json.loads(response_clean)
            except Exception:
                pass

        if isinstance(response_clean, str):
            if not (response_clean.startswith('{') or response_clean.startswith('[')):
                match = re.search(r'[\{\[].*[\}\]]', response_clean, re.DOTALL)
                if match:
                    response_clean = match.group()
            data = json.loads(response_clean)
        else:
            data = response_clean

        pattern = r'<x(\d+)><y(\d+)><x(\d+)><y(\d+)>'

        # 格式1: [{"bbox": [...]}] — 模型输出
        if isinstance(data, list):
            all_bbox_strs = []
            for item in data:
                all_bbox_strs.extend(item.get("bbox", []))
        # 格式2: {"bboxes": [...]} — GT
        elif isinstance(data, dict):
            all_bbox_strs = data.get("bboxes", [])
        else:
            return []

        bboxes = []
        for bbox_str in all_bbox_strs:
            match = re.search(pattern, bbox_str)
            if match:
                x1, y1, x2, y2 = map(int, match.groups())
                bboxes.append([x1, y1, x2, y2])
        return bboxes
    except Exception:
        return []


def load_conversation_data(data_file, max_samples=None):
    """加载对话格式的 .jsonl 数据"""
    data_list = []
    if not os.path.exists(data_file):
        print(f"错误: 找不到文件 {data_file}")
        return []

    with open(data_file, 'r', encoding='utf-8') as f:
        for idx, line in enumerate(f):
            if max_samples and idx >= max_samples:
                break
            try:
                data = json.loads(line.strip())
                convs = data['conversations']
                if isinstance(convs, str):
                    convs = json.loads(convs)
                
                human_val = next((c['value'] for c in convs if c['from'] == 'human'), None)
                gpt_val = next((c['value'] for c in convs if c['from'] == 'gpt'), None)
                
                if human_val:
                    data_list.append({
                        'image': data['image'],
                        'prompt': human_val,
                        'ground_truth': gpt_val or ""
                    })
            except Exception as e:
                print(f"解析第 {idx} 行失败: {e}")
                
    print(f"成功加载了 {len(data_list)} 条数据")
    return data_list


def run_vllm_inference(model_path, data_list, output_file, temperature=0.0):
    """核心推理函数"""
    print(f"\n正在初始化 vLLM 模型: {model_path}")
    
    # 加载 processor 用于生成正确的 prompt 格式
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    
    # 初始化 LLM 引擎
    llm = LLM(
        model=model_path,
        tensor_parallel_size=4,
        trust_remote_code=True,
        max_model_len=8192,
        gpu_memory_utilization=0.85,
        dtype="bfloat16",
        enforce_eager=True,
        limit_mm_per_prompt={"image": 1},
    )
    
    sampling_params = SamplingParams(
        temperature=temperature,
        top_p=0.9 if temperature > 0 else 1.0,
        max_tokens=2048,
        stop=["</s>", "<|im_end|>", "<|endoftext|>"],
        skip_special_tokens=False,
    )
    
    # 构建 vLLM 输入
    vllm_inputs = []
    for item in data_list:
        # 使用 processor 构建正确的消息格式
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": item['image']},
                    {"type": "text", "text": item['prompt']}
                ]
            }
        ]
        
        # 应用 chat template，生成带图片占位符的 prompt
        prompt_text = processor.apply_chat_template(
            messages, 
            tokenize=False, 
            add_generation_prompt=True
        )
        
        # 加载图片
        image = Image.open(item['image'])
        
        vllm_inputs.append({
            "prompt": prompt_text,
            "multi_modal_data": {"image": image}
        })
    
    print(f"开始执行 Batch 推理 (共 {len(data_list)} 个样本)...")
    print(f"Prompt 示例: {vllm_inputs[0]['prompt'][:200]}...")
    
    # 分批处理
    batch_size = 2
    all_results = []
    
    for i in range(0, len(vllm_inputs), batch_size):
        batch_inputs = vllm_inputs[i:i+batch_size]
        batch_data = data_list[i:i+batch_size]
        
        print(f"处理批次 {i//batch_size + 1}/{(len(vllm_inputs)-1)//batch_size + 1}")
        
        outputs = llm.generate(batch_inputs, sampling_params)
        
        for idx, (item, output) in enumerate(zip(batch_data, outputs)):
            res_text = output.outputs[0].text
            pred_boxes = extract_bboxes_from_response(res_text)
            gt_boxes = extract_bboxes_from_response(item['ground_truth'])
            
            all_results.append({
                'index': i + idx,
                'image': item['image'],
                'prompt': item['prompt'],
                'model_response': res_text,
                'ground_truth': item['ground_truth'],
                'pred_bboxes': pred_boxes,
                'gt_bboxes': gt_boxes,
                'num_pred': len(pred_boxes),
                'num_gt': len(gt_boxes),
            })
            
            if (i + idx + 1) % 5 == 0:
                print(f"进度: {i + idx + 1}/{len(data_list)}")
    
    # 保存完整结果
    with open(output_file, 'w', encoding='utf-8') as f:
        json.dump(all_results, f, ensure_ascii=False, indent=2)
    
    # 保存简化结果用于计算 IoU
    simplified_path = str(output_file).replace(".json", "_simplified.json")
    simplified = [{"image": r['image'], "gt_bboxes": r['gt_bboxes'], "pred_bboxes": r['pred_bboxes']} for r in all_results]
    with open(simplified_path, 'w', encoding='utf-8') as f:
        json.dump(simplified, f, ensure_ascii=False, indent=2)

    print(f"\n推理完成！")
    print(f"完整结果: {output_file}")
    print(f"简化结果: {simplified_path}")
    return all_results


def main():
    MODEL_PATH = "/home/disk2/hjl/ICL_QWEN/ckpt_0409_iter_26916"
    DATA_FILE = "/home/disk2/hjl/ICL_QWEN/ICL_benchmark/fewshot_data/conversation/conversation_k2.jsonl"
    OUTPUT_DIR = Path("/home/disk2/hjl/ICL_QWEN/new_test")
    OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

    IS_TEST = False
    
    if IS_TEST:
        print(">>> 进入测试模式 (Sample=10)")
        data = load_conversation_data(DATA_FILE, max_samples=10)
        output_name = "test_results_k2.json"
    else:
        print(">>> 进入全量推理模式")
        data = load_conversation_data(DATA_FILE)
        output_name = "full_results_k2.json"

    if data:
        run_vllm_inference(MODEL_PATH, data, OUTPUT_DIR / output_name, temperature=0.0)
    else:
        print("未加载到有效数据，请检查路径。")


if __name__ == "__main__":
    main()