RAGEN / cleanrl /README_convert.md
Harryis's picture
Add files using upload-large-folder tool
4c32b9f verified
|
Raw
History Blame Contribute Delete
3.49 kB

测试结果转换工具

这个工具可以将PPO训练过程中保存的测试结果从one-hot向量格式转换为更直观的10x10字母矩阵格式,方便LLM分析。

功能

  • 将observation中的one-hot编码转换回字母(A-E, X)
  • 重建10x10的网格地图
  • 将归一化的位置坐标转换回实际坐标
  • 在网格中标记agent的当前位置(用@表示)
  • 保留所有原始的统计信息和轨迹数据

使用方法

1. 转换单个文件

python convert_test_results.py /path/to/test_iter_732.json

这会在同一目录下生成 test_iter_732_readable.json

2. 指定输出文件

python convert_test_results.py /path/to/test_iter_732.json -o /path/to/output.json

3. 批量转换目录中的所有测试文件

python convert_test_results.py /path/to/runs/directory -a

这会递归查找所有 test_iter_*.json 文件并转换它们。

4. 自定义批量转换的文件模式

python convert_test_results.py /path/to/runs/directory -a -p "test_*.json"

示例

转换EASY难度的所有测试结果

python convert_test_results.py runs/Ultrahorizon-v0__ppo_ultrahorizon__1_EASY__1761810199 -a

转换特定的测试文件

python convert_test_results.py runs/Ultrahorizon-v0__ppo_ultrahorizon__1_EASY__1761810199/test_iter_732.json

输出格式

转换后的JSON文件包含:

{
  "iteration": 732,
  "global_step": 1497088,
  "training_progress": 0.1,
  "test_results": {
    "episode_returns": [...],
    "episode_lengths": [...],
    "final_scores": [...],
    "mean_return": 123.45,
    "trajectories": [
      {
        "actions": [0, 1, 2, ...],
        "rewards": [1.0, -0.5, ...],
        "dones": [false, false, ...],
        "observations": [
          {
            "step": 0,
            "position": {"x": 5, "y": 3},
            "energy": 20.0,
            "score": 0.0,
            "steps": 0.0,
            "grid": [
              ["A", "B", "C", ...],
              ["D", "E", "X", ...],
              ...
            ],
            "grid_string": "  0 1 2 3 4 5 6 7 8 9\n  -------------------\n0|A B C D E X A B C D \n1|E X A B C D E X A B \n...",
            "action": "right",
            "reward": 1.0
          },
          ...
        ]
      }
    ]
  }
}

网格可视化

每个observation都包含 grid_string 字段,显示10x10的网格:

  0 1 2 3 4 5 6 7 8 9
  -------------------
0|A B C D E X A B C D 
1|E X A B C D E X A B 
2|C D E X A B C D E X 
3|A B C D E @ A B C D   <- @ 表示agent位置
4|E X A B C D E X A B 
5|C D E X A B C D E X 
6|A B C D E X A B C D 
7|E X A B C D E X A B 
8|C D E X A B C D E X 
9|A B C D E X A B C D 

观察向量结构

原始observation是605维向量:

  • 位置 (2维): x, y 坐标,归一化到 [0, 1]
  • 能量 (1维): 归一化到 [0, 1],原始范围约为 [0, 20]
  • 分数 (1维): 当前累积分数
  • 步数 (1维): 归一化到 [0, 1],原始范围为 [0, 30]
  • 网格状态 (600维): 10x10网格,每个格子6维one-hot编码(A, B, C, D, E, X)

动作映射

  • 0: up (向上)
  • 1: down (向下)
  • 2: left (向左)
  • 3: right (向右)

注意事项

  1. 转换后的文件会比原始文件大,因为包含了可读的字符串表示
  2. 网格坐标系:y=0在顶部,y=9在底部
  3. agent位置用 @ 标记在grid_string中
  4. 所有原始数据(actions, rewards, dones)都会保留