def load_node(cls) -> NoReturn:
"""
加载节点目录中的key文件
:return:
"""
key_dir = PathUtils.join(GflConf.home_dir, "key")
cls.__load_node(PathUtils.join(key_dir, "key.json"))
def __init__(self, id):
super(DatasetPath, self).__init__(id)
self.__root_dir = PathUtils.join(GflConf.data_dir, "dataset", id)
self.__metadata_file = PathUtils.join(self.__root_dir, "metadata.json")
self.__config_dir = PathUtils.join(self.__root_dir, "dataset")
self.__dataset_config_file = PathUtils.join(self.__root_dir, "dataset",
"dataset.json")
self.__module_name = "fl_dataset"
self.__module_dir = PathUtils.join(self.__root_dir, "dataset")
def init_node(cls) -> NoReturn:
"""
initialize GFL node
"""
node = cls.__new_node()
key_dir = PathUtils.join(GflConf.home_dir, "key")
os.makedirs(key_dir, exist_ok=True)
key_file = PathUtils.join(key_dir, "key.json")
cls.__save_node(node, key_file)
cls.default_node = node
def init_node(cls) -> NoReturn:
"""
初始化GFL节点
:return:
"""
cls.__new_node()
key_dir = PathUtils.join(GflConf.home_dir, "key")
os.makedirs(key_dir, exist_ok=True)
key_file = PathUtils.join(key_dir, "key.json")
cls.__save_node(key_file)
def load_node(cls) -> NoReturn:
"""
Load the key file in the node directory, and create default_node and standalone_nodes objects
"""
key_dir = PathUtils.join(GflConf.home_dir, "key")
cls.default_node = cls.__load_node(PathUtils.join(key_dir, "key.json"))
for filename in os.listdir(key_dir):
if filename.startswith("node-"):
node_idx = int(filename[5:-5])
cls.standalone_nodes[node_idx] = cls.__load_node(
PathUtils.join(key_dir, filename))
def load_all_job(cls) -> List[Job]:
job_dir = PathUtils.join(GflConf.data_dir, "job")
jobs = []
for filename in os.listdir(job_dir):
path = PathUtils.join(job_dir, filename)
if os.path.isdir(path):
try:
job = cls.load_job(filename)
jobs.append(job)
except:
pass
return jobs
def load_node(cls) -> NoReturn:
"""
加载节点目录中的key文件，创建default_node和standalone_nodes对象
:return:
"""
key_dir = PathUtils.join(GflConf.home_dir, "key")
cls.default_node = cls.__load_node(PathUtils.join(key_dir, "key.json"))
for filename in os.listdir(key_dir):
if filename.startswith("node-"):
node_idx = int(filename[5:-5])
cls.standalone_nodes[node_idx] = cls.__load_node(
PathUtils.join(key_dir, filename))
def generate_config(cls, path: str = None) -> None:
"""
generate config file in ``path``.
:param path: the config file path, if it's None, will be replaced by './config.yaml'.
:return:
"""
if path is None:
path = "config.yaml"
src_path = PathUtils.join(PathUtils.src_root_dir(), "resources",
"config.yaml")
shutil.copy(src_path, path)
def load_logging_config(cls) -> None:
"""
"""
logging_config_path = PathUtils.join(PathUtils.src_root_dir(),
"resources", "logging.yaml")
with open(logging_config_path) as f:
text = f.read().replace("{logs_root}", GflConf.logs_dir)
data = yaml.load(text, yaml.SafeLoader)
if cls.get_property("debug"):
data["root"]["level"] = "DEBUG"
data["loggers"]["gfl"]["level"] = "DEBUG"
logging.config.dictConfig(data)
def reload(cls):
"""
Reload readonly parameters from the YAML file.
:return:
"""
with open(PathUtils.join(cls.__home_dir, "conf.yaml"), "r") as f:
cls.readonly_props = yaml.safe_load(f.read())
def save_job(cls, job: Job, *, module=None, module_data=None) -> NoReturn:
"""
Save job
:param job: job to save
:param module: job module
:param module_data:
"""
job_path = JobPath(job.job_id)
job_path.makedirs()
cls.__save_json(job_path.metadata_file, job.metadata)
cls.__save_json(job_path.job_config_file, job.job_config)
cls.__save_json(job_path.train_config_file, job.train_config)
cls.__save_json(job_path.aggregate_config_file, job.aggregate_config)
if module_data is not None:
ZipUtils.extract_data(module_data, GflConf.temp_dir)
ModuleUtils.migrate_module(
PathUtils.join(GflConf.temp_dir, job.job_id),
job_path.module_name, job_path.module_dir)
elif module is not None:
ModuleUtils.submit_module(module, job_path.module_name,
job_path.module_dir)
else:
ModuleUtils.submit_module(job.module, job_path.module_name,
job_path.module_dir)
def save_dataset(cls,
dataset: Dataset,
*,
module=None,
module_data=None) -> NoReturn:
"""
Save dataset
:param dataset: dataset to save
:param module: dataset module
"""
if module is None:
module = dataset.module
dataset_path = DatasetPath(dataset.dataset_id)
dataset_path.makedirs()
cls.__save_json(dataset_path.metadata_file, dataset.metadata)
cls.__save_json(dataset_path.dataset_config_file,
dataset.dataset_config)
if module_data is not None:
ZipUtils.extract_data(module_data, GflConf.temp_dir)
ModuleUtils.migrate_module(
PathUtils.join(GflConf.temp_dir, dataset.dataset_id),
dataset_path.module_name, dataset_path.module_dir)
elif module is not None:
ModuleUtils.submit_module(module, dataset_path.module_name,
dataset_path.module_dir)
else:
ModuleUtils.submit_module(dataset.module, dataset_path.module_name,
dataset_path.module_dir)
def init_node(cls) -> NoReturn:
"""
初始化GFL节点
:return:
"""
node = cls.__new_node()
key_dir = PathUtils.join(GflConf.home_dir, "key") # /Users/YY/.gfl/key
os.makedirs(key_dir, exist_ok=True)
key_file = PathUtils.join(key_dir,
"key.json") # /Users/YY/.gfl/key/key.json
cls.__save_node(node, key_file)
# key.json中的内容如下
# {
# "address": "a8C03cEBFc6C11C1707032590adf2ACF4ccAc655",
# "pub_key": "d2a95fb211c91f79d052c3c927f51b22893a3b3f7a28090f32d03fc7224bdca0be91173445f71bf1bf91d0fee52ee7c805b7b10dc1b12fa2ed5267b818eb1bc8",
# "priv_key": "708d8f67deb461bdf2a3c9c2d82584b8304cbad32398a5ce5706a8e45f5210bf"
# }
cls.default_node = node
def run(cls, role, console, **kwargs):
sys.stderr = open(os.devnull, "w")
cls.logger = logging.getLogger("gfl")
with Daemonizer() as (is_setup, daemonizer):
main_pid = None
if is_setup:
main_pid = os.getpid()
pid_file = PathUtils.join(GflConf.home_dir, "proc.lock")
stdout_file = PathUtils.join(GflConf.logs_dir, "console_out")
stderr_file = PathUtils.join(GflConf.logs_dir, "console_err")
is_parent = daemonizer(pid_file,
stdout_goto=stdout_file,
stderr_goto=stderr_file)
if is_parent:
if console and main_pid == os.getpid():
Shell.startup()
GflNode.load_node()
if GflConf.get_property("net.mode") == "standalone":
client_number = GflConf.get_property(
"net.standalone.client_number")
for _ in range(len(GflNode.standalone_nodes), client_number):
GflNode.add_standalone_node()
ManagerHolder.default_manager = NodeManager(
node=GflNode.default_node, role="server")
for i in range(client_number):
client_manager = NodeManager(node=GflNode.standalone_nodes[i],
role="client")
ManagerHolder.standalone_managers.append(client_manager)
else:
ManagerHolder.default_manager = NodeManager(
node=GflNode.default_node, role=role)
# cls.__startup_node_managers()
HttpListener.start()
while HttpListener.is_alive():
time.sleep(2)
class Log(LogBase):
debug_filename = PathUtils.join(GflConf.logs_dir, "debug.log")
info_filename = PathUtils.join(GflConf.logs_dir, "info.log")
warn_filename = PathUtils.join(GflConf.logs_dir, "warn.log")
error_filename = PathUtils.join(GflConf.logs_dir, "error.log")
def __init__(self, name):
super(Log, self).__init__(name)
def debug(self, msg, *params):
pass
def info(self, msg, *params):
pass
def warn(self, msg, *params):
pass
def error(self, msg, *params):
pass
def load(cls) -> None:
"""
load config properties from disk file.
:return:
"""
base_config_path = PathUtils.join(PathUtils.src_root_dir(),
"resources", "config.yaml")
with open(base_config_path) as f:
cls.__readonly_props = yaml.load(f, Loader=yaml.SafeLoader)
path = PathUtils.join(cls.home_dir, "config.yaml")
if os.path.exists(path):
with open(path) as f:
config_data = yaml.load(f, Loader=yaml.SafeLoader)
cls.__readonly_props.update(config_data)
if os.path.exists(cls.logs_dir):
cls.load_logging_config()
else:
warnings.warn("cannot found logs dir.")
def add_standalone_node(cls) -> NoReturn:
# 添加【一个】standalone_node
node = cls.__new_node()
for i in range(100):
# 限制最多100个模拟节点， 防止此处出现死循环
if i not in cls.standalone_nodes:
key_file = PathUtils.join(GflConf.home_dir, "key",
"node-%d.json" % i)
cls.__save_node(node, key_file)
cls.standalone_nodes[i] = node
return
raise ValueError("最多只支持100个standalone模式虚拟节点.")
def run(cls, console=True, **kwargs):
sys.stderr = open(os.devnull, "w")
cls.logger = logging.getLogger("gfl")
with Daemonizer() as (is_setup, daemonizer):
main_pid = None
if is_setup:
main_pid = os.getpid()
pid_file = PathUtils.join(GflConf.home_dir, "proc.lock")
stdout_file = PathUtils.join(GflConf.logs_dir, "console_out")
stderr_file = PathUtils.join(GflConf.logs_dir, "console_err")
is_parent = daemonizer(pid_file,
stdout_goto=stdout_file,
stderr_goto=stderr_file)
if is_parent:
if console and main_pid == os.getpid():
Shell.startup()
GflNode.load_node()
HttpListener.start()
NodeManager.get_instance().run()
def send_partial_params(cls, client: str, job_id: str, step: int,
params) -> NoReturn:
# 这里的参数client，暂时认为是client_address
# 在standalone模式下，trainer当前训练轮次得到的模型保存在指定路径下
client_params_dir = JobPath(job_id).client_params_dir(step, client)
os.makedirs(client_params_dir, exist_ok=True)
# 保存 job_id.pth为文件名
path = PathUtils.join(client_params_dir, job_id + '.pkl')
# path = client_params_dir + 'job_id.pth'
# torch.save(params, path)
with open(path, 'wb') as f:
pickle.dump(params, f)
print("训练完成，已将模型保存至：" + str(client_params_dir))
def run(cls, **kwargs):
daemon = kwargs.pop("daemon", False)
if daemon:
print("DAEMON")
with Daemonizer() as (is_setup, daemonizer):
if is_setup:
pass
pid_file = "proc.lock"
stdout_file = PathUtils.join(GflConf.logs_dir, "console_out")
stderr_file = PathUtils.join(GflConf.logs_dir, "console_err")
is_parent = daemonizer(pid_file,
stdout_goto=stdout_file,
stderr_goto=stderr_file)
if is_parent:
pass
GflConf.reload()
GflNode.load_node()
if GflConf.get_property("standalone.enabled"):
server_number = GflConf.get_property("standalone.server_number")
client_number = GflConf.get_property("standalone.client_number")
for _ in range(len(GflNode.standalone_nodes),
server_number + client_number):
GflNode.add_standalone_node()
for i in range(0, server_number):
node_manager = NodeManager(node=GflNode.standalone_nodes[i],
role="server")
cls.node_managers.append(node_manager)
for i in range(server_number, server_number + client_number):
node_manager = NodeManager(node=GflNode.standalone_nodes[i],
role="client")
cls.node_managers.append(node_manager)
else:
role = kwargs.pop("role")
print(role)
node_manager = NodeManager(node=GflNode.default_node, role=role)
cls.node_managers.append(node_manager)
cls.__startup_node_managers()
def add_standalone_node(cls) -> NoReturn:
"""
add standalone GFL node
"""
node = cls.__new_node()
for i in range(100):
# Limit up to 100 mock nodes to prevent an endless loop here
if i not in cls.standalone_nodes:
key_file = PathUtils.join(GflConf.home_dir, "key",
"node-%d.json" % i)
cls.__save_node(node, key_file)
cls.standalone_nodes[i] = node
return
raise ValueError("最多只支持100个standalone模式虚拟节点.")
def init(cls, force):
if os.path.exists(GflConf.home_dir):
if force:
logging.shutdown()
shutil.rmtree(GflConf.home_dir)
else:
raise ValueError("homedir not empty.")
# create home dir
os.makedirs(GflConf.home_dir)
# generate config file
GflConf.generate_config(PathUtils.join(GflConf.home_dir,
"config.yaml"))
# generate node address and key
GflNode.init_node()
# create data directories
Lfs.init()
def receive_global_params(cls, job_id: str, cur_round: int):
# 在standalone模式下，trainer获取当前聚合轮次下的全局模型
# 根据 Job 中的 job_id 和 cur_round 获取指定轮次聚合后的 全局模型参数的路径
global_params_dir = JobPath(job_id).global_params_dir(cur_round)
model_params_path = PathUtils.join(global_params_dir, job_id + '.pkl')
# 判断是否存在模型参数文件，如果存在则返回。
if os.path.exists(global_params_dir) and os.path.isfile(
model_params_path):
# resources_already:1
# self.__status = JobStatus.RESOURCE_ALREADY
print("训练方接收全局模型")
return model_params_path
else:
# 等待一段时间。在这段时间内获取到了模型参数文件，则返回
# 暂时不考虑这种情况
# 否则，认为当前模型参数文件已经无法获取
return None
def setUp(self) -> None:
self.dataset = generate_dataset()
print("dataset_id:" + self.dataset.dataset_id)
self.job = generate_job()
print("job_id:" + self.job.job_id)
self.job.mount_dataset(self.dataset)
GflNode.init_node()
node = GflNode.default_node
self.jobTrainerScheduler = JobTrainScheduler(node=node, job=self.job)
self.jobTrainerScheduler.register()
# aggregator需要初始化随机模型
global_params_dir = JobPath(self.job.job_id).global_params_dir(
self.job.cur_round)
# print("global_params_dir:"+global_params_dir)
os.makedirs(global_params_dir, exist_ok=True)
model_params_path = PathUtils.join(global_params_dir,
self.job.job_id + '.pth')
# print("model_params_path:"+model_params_path)
model = Net()
torch.save(model.state_dict(), model_params_path)
import logging.config
import os
import shutil
import tempfile
import warnings
import yaml
from gfl.utils import PathUtils
os_tempdir = tempfile.gettempdir()
gfl_tempdir = PathUtils.join(os_tempdir, "gfl")
if os.path.exists(gfl_tempdir):
os.makedirs(gfl_tempdir, exist_ok=True)
class GflConfMetadata(type):
@property
def home_dir(cls):
return cls._GflConf__home_dir
@home_dir.setter
def home_dir(cls, value):
cls._GflConf__home_dir = PathUtils.abspath(value)
cls._GflConf__data_dir = PathUtils.join(value, "data")
cls._GflConf__logs_dir = PathUtils.join(value, "logs")
cls._GflConf__cache_dir = PathUtils.join(value, "cache")
@property
def data_dir(cls):
return cls._GflConf__data_dir
def home_dir(cls, value):
cls._GflConf__home_dir = PathUtils.abspath(value)
cls._GflConf__data_dir = PathUtils.join(value, "data")
cls._GflConf__logs_dir = PathUtils.join(value, "logs")
cls._GflConf__cache_dir = PathUtils.join(value, "cache")
class GflConf(object, metaclass=GflConfMetadata):
# Parameters that can be modified at run time
__props = {}
# Parameters that are read from a configuration file and cannot be changed at run time
__readonly_props = {}
__home_dir = PathUtils.join(PathUtils.user_home_dir(), ".gfl")
__data_dir = PathUtils.join(__home_dir, "data")
__logs_dir = PathUtils.join(__home_dir, "logs")
__cache_dir = PathUtils.join(__home_dir, "cache")
__temp_dir = gfl_tempdir
@classmethod
def load(cls) -> None:
"""
load config properties from disk file.
:return:
"""
base_config_path = PathUtils.join(PathUtils.src_root_dir(),
"resources", "config.yaml")
with open(base_config_path) as f:
cls.__readonly_props = yaml.load(f, Loader=yaml.SafeLoader)
path = PathUtils.join(cls.home_dir, "config.yaml")
if os.path.exists(path):
with open(path) as f:
config_data = yaml.load(f, Loader=yaml.SafeLoader)
cls.__readonly_props.update(config_data)
if os.path.exists(cls.logs_dir):
cls.load_logging_config()
else:
warnings.warn("cannot found logs dir.")
@classmethod
def load_logging_config(cls) -> None:
"""
"""
logging_config_path = PathUtils.join(PathUtils.src_root_dir(),
"resources", "logging.yaml")
with open(logging_config_path) as f:
text = f.read().replace("{logs_root}", GflConf.logs_dir)
data = yaml.load(text, yaml.SafeLoader)
if cls.get_property("debug"):
data["root"]["level"] = "DEBUG"
data["loggers"]["gfl"]["level"] = "DEBUG"
logging.config.dictConfig(data)
@classmethod
def generate_config(cls, path: str = None) -> None:
"""
generate config file in ``path``.
:param path: the config file path, if it's None, will be replaced by './config.yaml'.
:return:
"""
if path is None:
path = "config.yaml"
src_path = PathUtils.join(PathUtils.src_root_dir(), "resources",
"config.yaml")
shutil.copy(src_path, path)
@classmethod
def set_config(cls, d: dict) -> None:
"""
Batch update config properties. Generally, this method is not recommend.
:param d: a dict represent config properties.
:return:
"""
cls.__props.update(d.copy())
@classmethod
def get_property(cls, key, default=None):
"""
Get the value of readonly parameters.
:param key: a string of the key to get the value
:param default: return value if key not found
"""
op_res, val = cls.__get_from_dict(cls.__props, cls.__split_key(key),
default)
if op_res:
return val
return cls.__get_from_dict(cls.__readonly_props, cls.__split_key(key),
default)[1]
@classmethod
def set_property(cls, key, value):
"""
Set parameters at run time.
:param key:
:param value:
:return:
"""
cls.__set_to_dict(cls.__props, cls.__split_key(key), value)
@classmethod
def remove_property(cls, key):
cls.__remove_from_dict(cls.__props, cls.__split_key(key))
@classmethod
def __split_key(cls, key: str):
if key is None or key.strip() == "":
raise ValueError("key cannot be none or empty.")
return key.split(".")
@classmethod
def __exists_in_dict(cls, d: dict, k_seq: list):
if k_seq is None or len(k_seq) == 0:
return False
for k in k_seq:
if k in d:
d = d[k]
else:
return False
return True
@classmethod
def __get_from_dict(cls, d: dict, k_seq: list, default=None):
if k_seq is None or len(k_seq) == 0:
raise ValueError("key cannot be none or empty")
for k in k_seq:
if k in d:
d = d[k]
else:
return False, default
return True, d
@classmethod
def __remove_from_dict(cls, d: dict, k_seq: list):
if k_seq is None or len(k_seq) == 0:
raise ValueError("key cannot be none or empty")
for k in k_seq[:-1]:
if k not in d:
return False
d = d[k]
try:
del d[k_seq[-1]]
return True
except:
return False
@classmethod
def __set_to_dict(cls, d: dict, k_seq: list, value):
if k_seq is None or len(k_seq) == 0:
raise ValueError("key cannot be none or empty")
for k in k_seq[:-1]:
if k not in d:
d[k] = {}
d = d[k]
d[k_seq[-1]] = value
def __load_json(cls, file_path, clazz):
if not PathUtils.exists(file_path):
return None
with open(file_path, "r") as f:
d = json.loads(f.read())
return clazz().from_dict(d)
def init(cls):
os.makedirs(GflConf.cache_dir)
os.makedirs(GflConf.data_dir)
os.makedirs(GflConf.logs_dir)
os.makedirs(PathUtils.join(GflConf.data_dir, "job"))
os.makedirs(PathUtils.join(GflConf.data_dir, "dataset"))
def __init__(self, id):
super(JobPath, self).__init__(id)
self.__root_dir = PathUtils.join(GflConf.data_dir, "job", id)
self.__metadata_file = PathUtils.join(self.__root_dir, "metadata.json")
self.__sqlite_file = PathUtils.join(self.__root_dir, "job.sqlite")
self.__config_dir = PathUtils.join(self.__root_dir, "job")
self.__job_config_file = PathUtils.join(self.__root_dir, "job",
"job.json")
self.__train_config_file = PathUtils.join(self.__root_dir, "job",
"train.json")
self.__aggregate_config_file = PathUtils.join(self.__root_dir, "job",
"aggregate.json")
self.__module_name = "fl_model"
self.__module_dir = PathUtils.join(self.__root_dir, "job")
self.__metrics_dir = PathUtils.join(self.__root_dir, "results",
"metrics")
self.__params_dir = PathUtils.join(self.__root_dir, "results",
"params")
self.__reports_dir = PathUtils.join(self.__root_dir, "results",
"reports")
self.__client_params_dir = PathUtils.join(self.__root_dir, "round-%d",
"%s", "params")
self.__client_word_dir = PathUtils.join(self.__root_dir, "round-%d",
"%s", "work")