eval/respace/tools/cot_gen_parallel_loop_v2.py

import json
import os
import time
from concurrent.futures import ThreadPoolExecutor, as_completed
from typing import List, Tuple
from openai import AzureOpenAI
from azure.identity import ChainedTokenCredential, AzureCliCredential, ManagedIdentityCredential, get_bearer_token_provider
import base64
import uuid
import re
from prompts import cot_gen_prompt_template_loop_v2
# from prompts import cot_gen_prompt_template_noloop_v2
import random

def encode_image_to_base64(image_path):
    """将图片文件编码为base64字符串"""
    if not os.path.exists(image_path):
        print(f"Warning: Image file not found at {image_path}")
        return None
    with open(image_path, "rb") as image_file:
        return base64.b64encode(image_file.read()).decode('utf-8')

def extract_json_from_response(content):
    """从消息内容中提取JSON，支持代码块格式和纯JSON格式"""
    try:
        # 方法1: 提取 ```json 代码块中的内容
        json_pattern = r'```json\s*\n?(.*?)\n?```'
        match = re.search(json_pattern, content, re.DOTALL)
        
        if match:
            json_str = match.group(1).strip()
        else:
            # 方法2: 如果没有代码块，查找第一个完整的JSON对象
            json_pattern_fallback = r'\{.*\}'
            match = re.search(json_pattern_fallback, content, re.DOTALL)
            
            if match:
                json_str = match.group(0).strip()
            else:
                raise ValueError("No valid JSON found in the message")
        
        # 尝试直接解析
        try:
            return json.loads(json_str)
        except json.JSONDecodeError:
            # 如果失败，尝试修复单引号问题
            fixed_json_str = json_str.replace("'", '"')
            try:
                return json.loads(fixed_json_str)
            except json.JSONDecodeError:
                try:
                    import ast
                    return ast.literal_eval(json_str)
                except (ValueError, SyntaxError):
                    raise ValueError(f"Unable to parse JSON: {json_str}")
        
    except Exception as e:
        if "JSON parsing error" in str(e):
            raise e
        else:
            raise ValueError(f"JSON parsing error: {e}")

def get_azure_client(instance_type='gcr'):
    """创建Azure OpenAI客户端"""
    scope = "api://trapi/.default"
    credential = get_bearer_token_provider(ChainedTokenCredential(
        AzureCliCredential(),
        ManagedIdentityCredential(),
    ), scope)

    api_version = '2024-12-01-preview'
    deployment_name = 'gpt-4o_2024-11-20'
    
    # 支持两个实例轮换使用
    if instance_type == 'gcr':
        instance = 'gcr/shared'
    else:
        instance = 'msra/shared'
    
    endpoint = f'https://trapi.research.microsoft.com/{instance}'

    client = AzureOpenAI(
        azure_endpoint=endpoint,
        azure_ad_token_provider=credential,
        api_version=api_version,
    )
    
    return client, deployment_name, instance

def get_pending_files(group_layout_path: str, output_path: str, user_input_path: str, render_img_path: str) -> List[Tuple[str, str, str, str, str]]:
    """获取待处理的文件列表"""
    pending_files = []
    
    for root, dirs, files in os.walk(group_layout_path):
        for file in files:
            if file.endswith('.json'):
                file_wo_json = file.replace('.json', '')
                output_file_path = os.path.join(output_path, file_wo_json + ".txt")
                
                # 检查是否已处理
                if os.path.exists(output_file_path):
                    continue
                
                layout_file_path = os.path.join(root, file)
                user_input_txt_path = os.path.join(user_input_path, file_wo_json + '.txt')
                
                # 检查用户输入文件是否存在
                if not os.path.exists(user_input_txt_path):
                    print(f"User input file not found for {file_wo_json + '.txt'}, skipping...")
                    continue
                
                # 检查图片是否存在
                diag_view_image = os.path.join(render_img_path, file_wo_json, "diag", 'frame.jpg')
                top_view_image = os.path.join(render_img_path, file_wo_json, "top", 'frame.jpg')
                
                if not os.path.exists(diag_view_image) or not os.path.exists(top_view_image):
                    print(f"Images not found for {file_wo_json}, skipping...")
                    continue
                
                pending_files.append((
                    layout_file_path,
                    user_input_txt_path,
                    diag_view_image,
                    top_view_image,
                    output_file_path
                ))
    
    return pending_files

def check_cot_quality(cot_content: str) -> bool:
    """检查CoT内容质量，确保至少有6个段落"""
    if not cot_content or not cot_content.strip():
        return False
    
    # 按段落分割（双换行或单换行）
    paragraphs = [p.strip() for p in cot_content.split('\n') if p.strip()]
    
    # 检查段落数量
    if len(paragraphs) < 6:
        return False
    
    # 检查总长度（至少应该有一定的内容）
    if len(cot_content.strip()) < 500:
        return False
    
    return True

def process_single_file(args: Tuple[Tuple[str, str, str, str, str], int]) -> Tuple[str, bool, str]:
    """处理单个文件"""
    file_data, worker_id = args
    layout_file_path, user_input_txt_path, diag_view_image, top_view_image, output_file_path = file_data
    filename = os.path.basename(layout_file_path)
    
    try:
        print(f"Worker {worker_id}: Processing {filename}")
        
        # 获取客户端（轮换使用两个实例）
        instance_type = 'gcr' if worker_id % 2 == 1 else 'msra'
        client, deployment_name, instance = get_azure_client(instance_type)
        print(f"Worker {worker_id}: Using {instance}")
        
        # 读取layout JSON
        with open(layout_file_path, 'r') as f:
            layout_json = json.load(f)
            layout_json_str = json.dumps(layout_json, indent=4)
        
        # 读取用户输入
        with open(user_input_txt_path, 'r') as f:
            user_input = f.read().strip()
        
        # 编码图片
        diag_base64 = encode_image_to_base64(diag_view_image)
        top_base64 = encode_image_to_base64(top_view_image)
        
        if not diag_base64 or not top_base64:
            return filename, False, "Failed to encode images"
        
        # 生成提示词文本
        # cot_prompt = random.choice([cot_gen_prompt_template_loop_v2, cot_gen_prompt_template_noloop_v2])
        cot_prompt = cot_gen_prompt_template_loop_v2
        prompt_text = cot_prompt.replace(
            "<<<TARGET_LAYOUT_JSON_HERE>>>", layout_json_str
        ).replace(
            "<<<DESIGN_BRIEF_HERE>>>", user_input
        )
        
        # API调用，增加重试机制和质量检查
        max_retries = 100
        max_quality_retries = 100  # 质量重试次数
        
        for quality_attempt in range(max_quality_retries + 1):
            cot_str = None
            
            # API调用重试
            for attempt in range(max_retries):
                try:
                    response = client.chat.completions.create(
                        model=deployment_name,
                        messages=[
                            {
                                "role": "user",
                                "content": [
                                    {"type": "text", "text": prompt_text},
                                    {
                                        "type": "image_url",
                                        "image_url": {
                                            "url": f"data:image/jpeg;base64,{diag_base64}",
                                            "detail": "high"
                                        }
                                    },
                                    {
                                        "type": "image_url",
                                        "image_url": {
                                            "url": f"data:image/jpeg;base64,{top_base64}",
                                            "detail": "high"
                                        }
                                    }
                                ]
                            }
                        ],
                        max_tokens=4096,
                        temperature=1.0,
                    )
                    
                    # 处理响应
                    cot_str = response.choices[0].message.content.replace('\n\n', '\n').strip()
                    break  # 成功则跳出重试循环
                    
                except Exception as e:
                    if attempt == max_retries - 1:
                        raise e
                    print(f"Worker {worker_id}: API retry {attempt + 1} for {filename}")
                    time.sleep(2 ** attempt)  # 指数退避
            
            # 检查CoT质量
            if cot_str and check_cot_quality(cot_str):
                break  # 质量合格，跳出质量重试循环
            else:
                if quality_attempt < max_quality_retries:
                    print(f"Worker {worker_id}: CoT quality insufficient for {filename}, retrying... (attempt {quality_attempt + 1})")
                    time.sleep(1)  # 短暂等待后重试
                else:
                    print(f"Worker {worker_id}: Warning - CoT quality still insufficient for {filename} after {max_quality_retries} retries, proceeding anyway")

        # 保存结果
        with open(output_file_path, 'w') as output_file:
            output_file.write(cot_str)
        
        # 最终质量检查报告
        quality_passed = check_cot_quality(cot_str)
        quality_status = "✅" if quality_passed else "⚠️"
        print(f"Worker {worker_id}: {quality_status} Successfully processed {filename} (Quality: {'Pass' if quality_passed else 'Warning'})")
        return filename, True, ""
        
    except Exception as e:
        error_msg = f"Worker {worker_id}: ❌ Error processing {filename}: {str(e)}"
        print(error_msg)
        return filename, False, str(e)

def process_layout_files_parallel(num_workers: int = 10):
    """并行处理layout files"""
    group_layout_path = "/home/v-meiszhang/amlt-project/respace/grouped_layouts_v2"
    render_img_path = "/home/v-meiszhang/amlt-project/respace/eval/viz/misc"
    user_input_path = "/home/v-meiszhang/amlt-project/respace/user_design_layouts_v3"
    output_path = "/home/v-meiszhang/amlt-project/respace/layout_design_cot_loop_v3.1"

    # 确保输出目录存在
    os.makedirs(output_path, exist_ok=True)
    
    # 获取待处理的文件
    pending_files = get_pending_files(group_layout_path, output_path, user_input_path, render_img_path)
    print(f"Found {len(pending_files)} files to process")
    
    if not pending_files:
        print("No files to process!")
        return
    
    # 准备任务参数
    task_args = [
        (file_data, i % num_workers + 1) 
        for i, file_data in enumerate(pending_files)
    ]
    
    # 统计变量
    successful_count = 0
    failed_count = 0
    failed_files = []
    
    print(f"🚀 Starting parallel processing with {num_workers} workers...")
    
    # 使用ThreadPoolExecutor进行并行处理
    with ThreadPoolExecutor(max_workers=num_workers) as executor:
        # 提交所有任务
        future_to_file = {
            executor.submit(process_single_file, args): args[0][0] 
            for args in task_args
        }
        
        # 收集结果
        for future in as_completed(future_to_file):
            file_path = future_to_file[future]
            try:
                filename, success, error_msg = future.result()
                if success:
                    successful_count += 1
                else:
                    failed_count += 1
                    failed_files.append(filename)
                    
            except Exception as exc:
                failed_count += 1
                filename = os.path.basename(file_path)
                failed_files.append(filename)
                print(f"Task for {filename} generated exception: {exc}")
    
    # 打印最终统计
    print("\n" + "="*60)
    print("🎉 Processing completed!")
    print(f"✅ Successfully processed: {successful_count} files")
    print(f"❌ Failed to process: {failed_count} files")
    print(f"📊 Total files: {len(pending_files)}")
    if len(pending_files) > 0:
        print(f"📈 Success rate: {successful_count/len(pending_files)*100:.1f}%")
    
    # 打印失败的文件
    if failed_files:
        print(f"\n❌ Failed files:")
        for filename in failed_files[:10]:  # 只显示前10个
            print(f"   - {filename}")
        if len(failed_files) > 10:
            print(f"   ... and {len(failed_files) - 10} more")

def process_layout_files_batch(batch_size: int = 50, num_workers: int = 10):
    """分批并行处理，避免一次性处理太多文件导致内存问题"""
    group_layout_path = "/home/v-meiszhang/amlt-project/respace/grouped_layouts_v2"
    render_img_path = "/home/v-meiszhang/amlt-project/respace/eval/viz/misc"
    user_input_path = "/home/v-meiszhang/amlt-project/respace/user_design_layouts_v3"
    output_path = "/home/v-meiszhang/amlt-project/respace/layout_design_cot_loop_v3.1"
    
    os.makedirs(output_path, exist_ok=True)
    
    # 获取待处理的文件
    pending_files = get_pending_files(group_layout_path, output_path, user_input_path, render_img_path)
    print(f"Found {len(pending_files)} files to process")
    
    if not pending_files:
        print("No files to process!")
        return
    
    # 分批处理
    total_successful = 0
    total_failed = 0
    batch_count = (len(pending_files) + batch_size - 1) // batch_size
    
    for batch_idx in range(0, len(pending_files), batch_size):
        batch_files = pending_files[batch_idx:batch_idx + batch_size]
        current_batch = batch_idx // batch_size + 1
        
        print(f"\n🔄 Processing batch {current_batch}/{batch_count} ({len(batch_files)} files)")
        
        # 准备当前批次的任务参数
        task_args = [
            (file_data, i % num_workers + 1) 
            for i, file_data in enumerate(batch_files)
        ]
        
        batch_successful = 0
        batch_failed = 0
        
        # 处理当前批次
        with ThreadPoolExecutor(max_workers=num_workers) as executor:
            future_to_file = {
                executor.submit(process_single_file, args): args[0][0] 
                for args in task_args
            }
            
            for future in as_completed(future_to_file):
                try:
                    filename, success, error_msg = future.result()
                    if success:
                        batch_successful += 1
                    else:
                        batch_failed += 1
                        
                except Exception as exc:
                    batch_failed += 1
                    print(f"Task exception: {exc}")
        
        total_successful += batch_successful
        total_failed += batch_failed
        
        print(f"Batch {current_batch} completed: ✅ {batch_successful} success, ❌ {batch_failed} failed")
        
        # 批次间短暂休息，避免API限流
        if current_batch < batch_count:
            print("Resting 2 seconds between batches...")
            time.sleep(2)
    
    # 最终统计
    print("\n" + "="*60)
    print("🎉 All batches completed!")
    print(f"✅ Total successful: {total_successful} files")
    print(f"❌ Total failed: {total_failed} files")
    print(f"📊 Total files: {len(pending_files)}")
    if len(pending_files) > 0:
        print(f"📈 Overall success rate: {total_successful/len(pending_files)*100:.1f}%")

def main():
    """主函数"""
    import argparse
    
    parser = argparse.ArgumentParser(description="Parallel CoT generation")
    parser.add_argument("--workers", "-w", type=int, default=20, 
                        help="Number of parallel workers (default: 10)")
    parser.add_argument("--batch-size", "-b", type=int, default=0,
                        help="Batch size for processing (0 = process all at once)")
    
    args = parser.parse_args()
    
    if args.batch_size > 0:
        print(f"🚀 Starting batch processing with {args.workers} workers, batch size: {args.batch_size}")
        process_layout_files_batch(args.batch_size, args.workers)
    else:
        print(f"🚀 Starting parallel processing with {args.workers} workers")
        process_layout_files_parallel(args.workers)

if __name__ == "__main__":
    main()