import os

import subprocess

import sys

import glob

import pandas as pd

import time

import numpy as np

# ================= 1. 实验配置区域 (在这里修改你的参数) =================

# 你想搜索的权重列表

WEIGHTS_TO_SEARCH = [0.01,0.02,0.03,0.04,0.05,0.06,0.07,0.08,0.09,0.1]

#WEIGHTS_TO_SEARCH = np.arange(0.00, 0.1 + 0.01, 0.01).tolist() # 从 0.00 到 0.10，步长 0.01

# 基础参数配置 (你想记录在 CSV 里的参数都写在这里)

# 脚本会自动将这些参数拼接成命令行，并记录到结果表中

BASE_CONFIG = {

"dataset": "mosei",

#"common_dim": 256, # 你想调整的维度

#"layers": 2, # Transformer 层数

"batch_size": 128, # Batch Size

#"num_epochs": 50, # 训练轮数

"lr_main": 8e-5, # 学习率

"clip": 5.0, # 梯度裁剪阈值

"common_dim": 128, # 公共空间维度

#"dropout_prj": 0.5, # 投影层 Dropout 比例

#"attn_dropout": 0.2, # 注意力 Dropout 比例

}

# 结果保存文件名

SUMMARY_FILE = "tuning_summary.csv"

# Python解释器

PYTHON_EXE = sys.executable

# ======================================================================

def get_latest_log_csv(dataset_name):

"""找到 logs/dataset/ 目录下最新的 metrics_xxx.csv"""

log_dir = os.path.join("logs", dataset_name)

if not os.path.exists(log_dir):

return None

# 获取所有 metrics_*.csv 文件

list_of_files = glob.glob(os.path.join(log_dir, "metrics_*.csv"))

if not list_of_files:

return None

# 按创建时间排序，找最新的

latest_file = max(list_of_files, key=os.path.getctime)

return latest_file

def analyze_log_file(csv_path):

"""读取单次训练的 CSV，找到 Acc-2 最高的那一行"""

try:

# 尝试读取，如果文件为空或格式错误，Pandas 会报错

df = pd.read_csv(csv_path)

if df.empty or 'Test_Acc_2' not in df.columns:

print(f"⚠️ 日志文件 {csv_path} 为空或缺少关键列。")

return None

# 找到 Test_Acc_2 最大值对应的索引

best_idx = df['Test_Acc_2'].idxmax()

best_row = df.iloc[best_idx]

return {

"Best_Epoch": int(best_row['Epoch']),

"Best_Acc2": best_row['Test_Acc_2'],

"Best_F1": best_row['Test_F1'],

"MAE": best_row['Test_MAE'],

"Corr": best_row['Test_Corr']

}

except pd.errors.EmptyDataError:

print(f"⚠️ 日志文件 {csv_path} 是空的。")

return None

except Exception as e:

print(f"❌ 读取日志文件出错 {csv_path}: {e}")

return None

def main():

# 1. 准备 CSV 表头

config_keys = list(BASE_CONFIG.keys())

result_keys = ["Best_Acc2", "Best_F1", "MAE", "Corr", "Best_Epoch"]

# 强制定义列的顺序

headers = ["Timestamp", "Aux_Weight"] + config_keys + result_keys + ["Log_File"]

# 初始化汇总文件

if not os.path.exists(SUMMARY_FILE):

pd.DataFrame(columns=headers).to_csv(SUMMARY_FILE, index=False)

print(f"🚀 开始自动实验。共 {len(WEIGHTS_TO_SEARCH)} 组参数。")

print(f"📂 结果将保存在: {SUMMARY_FILE}\n")

for weight in WEIGHTS_TO_SEARCH:

print(f"==================================================")

print(f"▶ 正在运行: Aux_Weight = {weight}, Config = {BASE_CONFIG}")

print(f"==================================================")

# 2. 动态构造命令

# 基础命令

cmd = [PYTHON_EXE, "-m", "src.main", "--aux_weight", str(weight)]

# 将 BASE_CONFIG 里的键值对拼接到命令中

# 例如: --common_dim 256

for key, value in BASE_CONFIG.items():

cmd.append(f"--{key}")

cmd.append(str(value))

# 打印完整命令供检查

print(f"执行命令: {' '.join(cmd)}")

# 3. 运行训练

try:

# check=True 确保如果 python main.py 报错，脚本能捕获并提示

subprocess.run(cmd, check=True)

except subprocess.CalledProcessError:

print(f"❌ 实验 (Weight={weight}) 运行失败（Exit Code != 0）！跳过...")

continue

except KeyboardInterrupt:

print("\n⛔ 用户手动中断。")

break

# 4. 获取日志

latest_csv = get_latest_log_csv(BASE_CONFIG['dataset'])

if not latest_csv:

print("⚠️ 未找到日志文件（可能是训练未启动或目录被删），跳过记录。")

continue

# 5. 分析结果

result = analyze_log_file(latest_csv)

if result:

print(f"✅ 实验成功! Best Acc: {result['Best_Acc2']:.4f}")

# 6. 构造汇总行数据

row_data = {

"Timestamp": time.strftime('%Y-%m-%d %H:%M:%S'),

"Aux_Weight": weight,

**BASE_CONFIG,

**result,

"Log_File": latest_csv

}

# ================= [关键修改在这里] =================

# 创建 DataFrame 时，显式传入 columns=headers

# 这样 Pandas 就会强制按照 headers 的顺序排列数据，绝不会错位

df_row = pd.DataFrame([row_data], columns=headers)

df_row.to_csv(SUMMARY_FILE, mode='a', header=False, index=False)

# ==================================================

else:

print("⚠️ 无法从日志中提取有效结果。")

time.sleep(1)

print(f"\n🎉 所有实验结束！请打开 {SUMMARY_FILE} 查看对比结果。")

if __name__ == "__main__":

main()