analysis/decoupling_analysis/2loss/extract_training_data.py

#!/usr/bin/env python3
"""
从训练日志中提取loss数据，保存为JSONL格式。

2loss 实验说明：
- 去掉了 region loss
- context loss 权重调整为 0.25（和 DeCLIP 一致）
- 因此不需要进行归一化，直接对比
"""

import re
import json
import os
from pathlib import Path


# 配置
DECLIP_LOG = "/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/logs/DeCLIP_EVA-B_DINOv2-B_560/out.log"
INTEGRATED_2LOSS_LOG = "/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/logs/Integrated_EVA-B_DINOv2-B_560_2loss/out.log"
OUTPUT_DIR = "/mnt/bn/strategy-mllm-train/user/wangjunjie/code/xiaomoguhzz/DeCLIP_private/decoupling_analysis/2loss/results/data"


def parse_training_line(line: str) -> dict | None:
    """
    解析训练日志行，提取loss数据。
    
    格式示例:
    2026-01-22,15:50:21 | INFO | Train Epoch: 0 [    16/118287 (0%)] ... Loss_context: 2.2152 (2.2152) Loss_content: 0.55079 (0.55079) Loss: 2.7660 (2.7660)
    """
    if "Train Epoch:" not in line:
        return None
    
    try:
        # 提取时间戳
        timestamp_match = re.match(r'(\d{4}-\d{2}-\d{2},\d{2}:\d{2}:\d{2})', line)
        timestamp = timestamp_match.group(1) if timestamp_match else None
        
        # 提取epoch
        epoch_match = re.search(r'Train Epoch:\s*(\d+)', line)
        epoch = int(epoch_match.group(1)) if epoch_match else None
        
        # 提取当前step和总step
        step_match = re.search(r'\[\s*(\d+)/(\d+)', line)
        current_step = int(step_match.group(1)) if step_match else None
        total_steps = int(step_match.group(2)) if step_match else None
        
        # 提取LR
        lr_match = re.search(r'LR:\s*([\d.e+-]+)', line)
        lr = float(lr_match.group(1)) if lr_match else None
        
        # 提取Loss_context (当前值)
        context_match = re.search(r'Loss_context:\s*([\d.e+-]+)', line)
        loss_context = float(context_match.group(1)) if context_match else None
        
        # 提取Loss_content (当前值)
        content_match = re.search(r'Loss_content:\s*([\d.e+-]+)', line)
        loss_content = float(content_match.group(1)) if content_match else None
        
        # 提取总Loss (当前值) - 注意匹配 "Loss:" 但不匹配 "Loss_xxx:"
        total_match = re.search(r'Loss:\s*([\d.e+-]+)\s*\(', line)
        loss_total = float(total_match.group(1)) if total_match else None
        
        if all(v is not None for v in [epoch, loss_context, loss_content, loss_total]):
            return {
                "timestamp": timestamp,
                "epoch": epoch,
                "step": current_step,
                "total_steps": total_steps,
                "lr": lr,
                "loss_context": loss_context,
                "loss_content": loss_content,
                "loss_total": loss_total
            }
    except Exception as e:
        print(f"解析行失败: {line[:100]}... 错误: {e}")
    
    return None


def parse_validation_line(line: str) -> dict | None:
    """
    解析验证指标行。
    """
    if "rois.thing.macc1" not in line:
        return None
    
    try:
        # 提取时间戳
        timestamp_match = re.match(r'(\d{4}-\d{2}-\d{2},\d{2}:\d{2}:\d{2})', line)
        timestamp = timestamp_match.group(1) if timestamp_match else None
        
        # 提取字典部分
        dict_match = re.search(r'\{[^}]+\}', line)
        if dict_match:
            dict_str = dict_match.group(0).replace("'", '"')
            metrics = json.loads(dict_str)
            
            return {
                "timestamp": timestamp,
                "metrics": metrics
            }
    except Exception as e:
        print(f"解析验证行失败: {line[:100]}... 错误: {e}")
    
    return None


def extract_from_log(log_path: str, model_name: str) -> tuple[list, list]:
    """
    从日志文件中提取训练loss和验证指标。
    """
    training_data = []
    validation_data = []
    
    print(f"正在解析日志: {log_path}")
    
    with open(log_path, 'r', encoding='utf-8') as f:
        for line in f:
            # 尝试解析训练行
            train_record = parse_training_line(line)
            if train_record:
                train_record["model"] = model_name
                training_data.append(train_record)
                continue
            
            # 尝试解析验证行
            val_record = parse_validation_line(line)
            if val_record:
                val_record["model"] = model_name
                validation_data.append(val_record)
    
    print(f"  - 提取到 {len(training_data)} 条训练记录")
    print(f"  - 提取到 {len(validation_data)} 条验证记录")
    
    return training_data, validation_data


def save_jsonl(data: list, filepath: str):
    """保存数据为JSONL格式。"""
    os.makedirs(os.path.dirname(filepath), exist_ok=True)
    with open(filepath, 'w', encoding='utf-8') as f:
        for record in data:
            f.write(json.dumps(record, ensure_ascii=False) + '\n')
    print(f"已保存: {filepath} ({len(data)} 条记录)")


def main():
    print("=" * 60)
    print("训练数据提取工具 (2loss 版本)")
    print("=" * 60)
    print("说明: 2loss 实验已将 context loss 权重调整为 0.25，无需归一化")
    print()
    
    # 提取DeCLIP数据
    declip_train, declip_val = extract_from_log(DECLIP_LOG, model_name="DeCLIP")
    
    # 提取Integrated_2loss数据
    integrated_train, integrated_val = extract_from_log(
        INTEGRATED_2LOSS_LOG,
        model_name="Integrated_2loss"
    )
    
    print()
    print("=" * 60)
    print("保存数据")
    print("=" * 60)
    
    # 保存训练数据
    save_jsonl(declip_train, os.path.join(OUTPUT_DIR, "declip_training.jsonl"))
    save_jsonl(integrated_train, os.path.join(OUTPUT_DIR, "integrated_2loss_training.jsonl"))
    
    # 保存验证数据
    all_validation = declip_val + integrated_val
    save_jsonl(all_validation, os.path.join(OUTPUT_DIR, "validation_metrics.jsonl"))
    
    # 打印统计信息
    print()
    print("=" * 60)
    print("数据统计")
    print("=" * 60)
    
    print("\n【DeCLIP训练数据】")
    if declip_train:
        epochs = set(r["epoch"] for r in declip_train)
        print(f"  Epoch范围: {min(epochs)} - {max(epochs)}")
        print(f"  总记录数: {len(declip_train)}")
        print(f"  最终Loss: context={declip_train[-1]['loss_context']:.4f}, "
              f"content={declip_train[-1]['loss_content']:.4f}, "
              f"total={declip_train[-1]['loss_total']:.4f}")
    
    print("\n【Integrated_2loss训练数据】")
    if integrated_train:
        epochs = set(r["epoch"] for r in integrated_train)
        print(f"  Epoch范围: {min(epochs)} - {max(epochs)}")
        print(f"  总记录数: {len(integrated_train)}")
        print(f"  最终Loss: context={integrated_train[-1]['loss_context']:.4f}, "
              f"content={integrated_train[-1]['loss_content']:.4f}, "
              f"total={integrated_train[-1]['loss_total']:.4f}")
    
    print("\n【验证指标】")
    for val in all_validation:
        print(f"  {val['model']} @ {val['timestamp']}:")
        m = val['metrics']
        print(f"    rois.thing.macc1: {m['rois.thing.macc1']:.4f}")
        print(f"    maskpool.thing.macc1: {m['maskpool.thing.macc1']:.4f}")
    
    print()
    print("=" * 60)
    print("提取完成!")
    print("=" * 60)


if __name__ == "__main__":
    main()