究极小白详解pymarl框架的main.py函数

赛博炼丹术士

已于 2023-12-08 14:58:46 修改

阅读量825

点赞数 13

文章标签：人工智能学习 python pygame 深度学习

于 2023-10-02 01:42:44 首次发布

本文链接：https://blog.csdn.net/m0_62313824/article/details/133464131

版权

ex.run_commandline(params)

sys与fd

SETTINGS['CAPTURE_MODE'] = "sys" # set to "no" if you want to see stdout/stderr in console

此处为设置捕获模式，其实就是设定是什么系统，如果是windows系统，就用"sys"，其它系统就要用"fd"。源码如下：

SETTINGS = FrozenKeyMunch.fromDict(
    {
        "CONFIG": {
            # make sure all config keys are compatible with MongoDB
            "ENFORCE_KEYS_MONGO_COMPATIBLE": True,
            # make sure all config keys are serializable with jsonpickle
            # THIS IS IMPORTANT. Only deactivate if you know what you're doing.
            "ENFORCE_KEYS_JSONPICKLE_COMPATIBLE": True,
            # make sure all config keys are valid python identifiers
            "ENFORCE_VALID_PYTHON_IDENTIFIER_KEYS": False,
            # make sure all config keys are strings
            "ENFORCE_STRING_KEYS": False,
            # make sure no config key contains an equals sign
            "ENFORCE_KEYS_NO_EQUALS": True,
            # if true, all dicts and lists in the configuration of a captured
            # function are replaced with a read-only container that raises an
            # Exception if it is attempted to write to those containers
            "READ_ONLY_CONFIG": True,
            # regex patterns to filter out certain IDE or linter directives
            # from inline comments in the documentation
            "IGNORED_COMMENTS": ["^pylint:", "^noinspection"],
            # if true uses the numpy legacy API, i.e. _rnd in captured functions is
            # a numpy.random.RandomState rather than numpy.random.Generator.
            # numpy.random.RandomState became legacy with numpy v1.19.
            "NUMPY_RANDOM_LEGACY_API": version.parse(opt.np.__version__)
            < version.parse("1.19")
            if opt.has_numpy
            else False,
        },
        "HOST_INFO": {
            # Collect information about GPUs using the nvidia-smi tool
            "INCLUDE_GPU_INFO": True,
            # Collect information about CPUs using py-cpuinfo
            "INCLUDE_CPU_INFO": True,
            # List of ENVIRONMENT variables to store in host-info
            "CAPTURED_ENV": [],
        },
        "COMMAND_LINE": {
            # disallow string fallback, if parsing a value from command-line
            # failed
            "STRICT_PARSING": False,
            # show command line options that are disabled (e.g. unmet
            # dependencies)
            "SHOW_DISABLED_OPTIONS": True,
        },
        # configure how stdout/stderr are captured. ['no', 'sys', 'fd']
        "CAPTURE_MODE": "sys" if platform.system() == "Windows" else "fd",
        # configure how dependencies are discovered. [none, imported, sys, pkg]
        "DISCOVER_DEPENDENCIES": "imported",
        # configure how source-files are discovered. [none, imported, sys, dir]
        "DISCOVER_SOURCES": "imported",
        # Configure the default beat interval, in seconds
        "DEFAULT_BEAT_INTERVAL": 10.0,
    },
)

其中，源码中的此键值对表明windows与其他系统区别

"CAPTURE_MODE": "sys" if platform.system() == "Windows" else "fd",

实例化

logger = get_logger()
ex = Experiment("yaml")
ex.logger = logger

获得logger方法，将Experiment实例化，并将logger方法赋给ex实例。

ex.captured_out_filter = apply_backspaces_and_linefeeds

设置输出文件的格式，避免有些实时输出（进度条等）不适合文件输出的形式

results_path = os.path.join(dirname(dirname(abspath(__file__))), "results")

dirname(abspath(__file__))是获得当前执行脚本的路径，abspath(__file__)是获得其绝对路径。如main.py在E:\fuxian\CADP-main\CADP-VD\src，则dirname(abspath(__file__))得输出即为：E:\fuxian\CADP-main\CADP-VD\src。dirname(dirname(abspath(__file__)))输出为上一条路径，为：E:\fuxian\CADP-main\CADP-VD，最后将results加入路径得到结果：results_path =E:\fuxian\CADP-main\CADP-VD\results。

params

if __name__ == '__main__':
    params = deepcopy(sys.argv)

此时的params值为[E:\fuxian\CADP-main\CADP-VD\src\main.py]，即只有脚本路径一个值的列表。

python src/main.py --config=qmix --env-config=sc2 with env_args.map_name=2s3z

在终端的命令语句如上，其实意思是将"--config=QMIX_CADP"，"--env-config=sc2"俩个值加入params这个列表中。最终params值为：[E:\fuxian\CADP-main\CADP-VD\src\main.py,"--config=QMIX_CADP","--env-config=sc2"]。其实也就相当于用append放入main.py函数中，如下：

    params.append("--config=QMIX_CADP")
    params.append("--env-config=sc2")

with env_args.map_name=2s3z

等同于将map_name=2s3z放入env,yaml文件中的env_args参数项下。

config_ dict

with open(os.path.join(os.path.dirname(__file__), "config", "default.yaml"), "r") as f:
        try:
            config_dict = yaml.load(f, Loader=yaml.FullLoader)
        except yaml.YAMLError as exc:
            assert False, "default.yaml error: {}".format(exc)

打开E:\fuxian\CADP-main\CADP-VD\src\config\default.yaml文件，并将里面的参数作为字典传给config_dict。

    # Load algorithm and env base configs
    env_config = _get_config(params, "--env-config", "envs")
    alg_config = _get_config(params, "--config", "algs")

def _get_config(params, arg_name, subfolder):
    config_name = None
    for _i, _v in enumerate(params):
        if _v.split("=")[0] == arg_name:
            config_name = _v.split("=")[1]
            del params[_i]
            break

    if config_name is not None:
        with open(os.path.join(os.path.dirname(__file__), "config", subfolder, "{}.yaml".format(config_name)), "r") as f:
            try:
                config_dict = yaml.load(f, Loader=yaml.FullLoader)
            except yaml.YAMLError as exc:
                assert False, "{}.yaml error: {}".format(config_name, exc)
        return config_dict

检查params中是否有参数--env-config，如果有，就将等号后面的"sc2"传递给函数内的参数config_name。进入if判断后，打开E:\fuxian\CADP-main\CADP-VD\src\config\envs\sc2.yaml文件，将里面的参数作为字典，传给env_config。alg_config同。

# config_dict = {**config_dict, **env_config, **alg_config}
    config_dict = recursive_dict_update(config_dict, env_config)
    config_dict = recursive_dict_update(config_dict, alg_config)

def recursive_dict_update(d, u):
    for k, v in u.items():
        # 两个参数表示的意思为：
        # object – 实例对象，就相当于刚才例子中的2.
        # classinfo – 可以是直接或间接类名、基本类型或者由它们组成的元组。
        # 返回值：如果对象的类型与参数二的类型（classinfo）相同则返回 True，否则返回 False。
        new_var = collections.abc.Mapping
        if isinstance(v, new_var):
            # 通过key值获取字典内容并返回，如果没有key可以返回指定值
            d[k] = recursive_dict_update(d.get(k, {}), v)
        else:
            d[k] = v
    return d

更新config_dict，将env_config这个字典作为值，“env_config”作为键，更新并入config_dict字典中。

# now add all the config to sacred
    ex.add_config(config_dict)

将config_dict字典作为参数配置，传给ex实例中。

logger

# Save to disk by default for sacred
    logger.info("Saving to FileStorageObserver in results/sacred.")
    file_obs_path = os.path.join(results_path, "sacred")
    ex.observers.append(FileStorageObserver.create(file_obs_path))

在终端输出一句等级为info的话"Saving to FileStorageObserver in results/sacred."

创建file_obs_path=E:\fuxian\CADP-main\CADP-VD\src\sacred

创建一个ex的观察者文件(其实就是写日志的)，存储路径为file_obs_path。

ex.run_commandline(params)

 ex.run_commandline(params)

@ex.main
def my_main(_run, _config, _log):
    # Setting the random seed throughout the modules
    config = config_copy(_config)
    np.random.seed(config["seed"])
    th.manual_seed(config["seed"])
    config['env_args']['seed'] = config["seed"]

    # run the framework
    run(_run, config, _log)

ex.run_commandline(params)语句将ex中的参数配置，logger等信息汇总传给my_main函数，最后运行my_main函数，并且自动进行各种关键信息的输出，

@ex.main是一个装饰器，the @ex.main is just a short way of saying:my_main=ex.main(my_main)。相当于把my_main函数放进ex类的main方法中。最后的ex.run_commandline(params)语句也会调用此函数。

在my_main函数中，主要设置配置参数，随机数种子等。最后的run函数为训练等一系列函数。

具体ex.run_commandline(params)这行命令怎么调用的my_main函数，我debug了一下午，倒是有些眉目，但是里面给中嵌套调来调去，各种@装饰来装饰去，及其复杂，研究明白了可能会写ex.run_commandline(params)函数调用my_main详解。也可能因为过于复杂而不写。挖个坑吧。如果大家需要叫多，可留言。

（刚入坑多智能体的菜鸡博士一枚，请多多指教。