AWS Sagemaker使用SM_USER_ARGS
(已记录here)作为环境变量,其中包含用户传递的自变量的字符串(列表)。因此,环境变量值如下所示:'["--test_size","0.2","--random_seed","42", "--not_optmize"]'
。
使用json.loads()
,我能够将该字符串转换为python列表。虽然,我想创建一个抽象模块,以返回argparse命名空间的方式,无论我在本地还是在AWS Sagemaker服务中运行其余代码,该代码都保持不变。
所以,基本上,我想要的是一个接收输入["--test_size","0.2","--random_seed","42", "--not_optmize"]
和输出Namespace(test_size=0.2, random_seed='42', not_optmize=True, <other_arguments>... ])
的代码。
python argparse包可以帮助我吗?我正在尝试找出不需要重新实现argparse解析器的方法。
这里是一个例子,我有这个config.ini文件:
[Docker]
home_dir = /opt
SM_MODEL_DIR = %(home_dir)s/ml/model
SM_CHANNELS = ["training"]
SM_NUM_GPUS = 1
SM_NUM_CPUS =
SM_LOG_LEVEL = 20
SM_USER_ARGS = ["--test_size","0.2","--random_seed","42"]
SM_INPUT_DIR = %(home_dir)s/ml/input
SM_INPUT_CONFIG_DIR = %(home_dir)s/ml/input/config
SM_OUTPUT_DIR = %(home_dir)s/ml/output
SM_OUTPUT_INTERMEDIATE_DIR = %(home_dir)s/ml/output/intermediate
我有这个Argparser类:
import argparse
import configparser
import datetime
import json
import multiprocessing
import os
import time
from pathlib import Path
from typing import Any, Dict
from .files import JsonFile, YAMLFile
class ArgParser(ABC):
@abstractmethod
def get_arguments(self) -> Dict[str, Any]:
pass
class AWSArgParser(ArgParser):
def __init__(self):
configuration_file_path = 'config.ini'
self.environment = "Sagemaker" \
if os.environ.get("SM_MODEL_DIR", False) \
else os.environ.get("ENVIRON", "Default")
config = configparser.ConfigParser()
config.read(configuration_file_path)
if self.environment == "Local":
config[self.environment]["home_dir"] = str(pathlib.Path(__file__).parent.absolute())
if self.environment != 'Sagemaker':
config[self.environment]["SM_NUM_CPUS"] = str(multiprocessing.cpu_count())
for key, value in config[self.environment].items():
os.environ[key.upper()] = value
self.parser = argparse.ArgumentParser()
# AWS Sagemaker default environmental arguments
self.parser.add_argument(
'--model_dir',
type=str,
default=os.environ['SM_MODEL_DIR'],
)
self.parser.add_argument(
'--channel_names',
default=json.loads(os.environ['SM_CHANNELS']),
)
self.parser.add_argument(
'--num_gpus',
type=int,
default=os.environ['SM_NUM_GPUS'],
)
self.parser.add_argument(
'--num_cpus',
type=int,
default=os.environ['SM_NUM_CPUS'],
)
self.parser.add_argument(
'--user_args',
default=json.loads(os.environ['SM_USER_ARGS']),
)
self.parser.add_argument(
'--input_dir',
type=str,
default=os.environ['SM_INPUT_DIR'],
)
self.parser.add_argument(
'--input_config_dir',
type=Path,
default=os.environ['SM_INPUT_CONFIG_DIR'],
)
self.parser.add_argument(
'--output_dir',
type=Path,
default=os.environ['SM_OUTPUT_DIR'],
)
# Extra arguments
self.run_tag = datetime.datetime \
.fromtimestamp(time.time()) \
.strftime('%Y-%m-%d-%H-%M-%S')
self.parser.add_argument(
'--run_tag',
default=self.run_tag,
type=str,
help=f"Run tag (default: 'datetime.fromtimestamp')",
)
self.parser.add_argument(
'--environment',
type=str,
default=self.environment,
)
self.args = self.parser.parse_args()
def get_arguments(self) -> Dict[str, Any]:
<parse self.args.user_args>
return self.args
然后我有了我的train
脚本:
from utils.arg_parser import AWSArgParser
if __name__ == '__main__':
logger.info(f"Begin train.py")
if os.environ["ENVIRON"] == "Sagemaker":
arg_parser = AWSArgParser()
args = arg_parser.get_arguments()
else:
args = <normal local parse>
在@chepner的注释之后,示例解决方案将是这样的:
config.ini文件:
[Docker]
home_dir = /opt
SM_MODEL_DIR = %(home_dir)s/ml/model
SM_CHANNELS = ["training"]
SM_NUM_GPUS = 1
SM_NUM_CPUS =
SM_LOG_LEVEL = 20
SM_USER_ARGS = ["--test_size","0.2","--random_seed","42", "--not_optimize"]
SM_INPUT_DIR = %(home_dir)s/ml/input
SM_INPUT_CONFIG_DIR = %(home_dir)s/ml/input/config
SM_OUTPUT_DIR = %(home_dir)s/ml/output
SM_OUTPUT_INTERMEDIATE_DIR = %(home_dir)s/ml/output/intermediate
TrainArgParser
类,例如:
class ArgParser(ABC):
@abstractmethod
def get_arguments(self) -> Dict[str, Any]:
pass
class TrainArgParser(ArgParser):
def __init__(self):
configuration_file_path = 'config.ini'
self.environment = "Sagemaker" \
if os.environ.get("SM_MODEL_DIR", False) \
else os.environ.get("ENVIRON", "Default")
config = configparser.ConfigParser()
config.read(configuration_file_path)
if self.environment == "Local":
config[self.environment]["home_dir"] = str(pathlib.Path(__file__).parent.absolute())
if self.environment != 'Sagemaker':
config[self.environment]["SM_NUM_CPUS"] = str(multiprocessing.cpu_count())
for key, value in config[self.environment].items():
os.environ[key.upper()] = value
self.parser = argparse.ArgumentParser()
# AWS Sagemaker default environmental arguments
self.parser.add_argument(
'--model_dir',
type=str,
default=os.environ['SM_MODEL_DIR'],
)
self.parser.add_argument(
'--channel_names',
default=json.loads(os.environ['SM_CHANNELS']),
)
self.parser.add_argument(
'--num_gpus',
type=int,
default=os.environ['SM_NUM_GPUS'],
)
self.parser.add_argument(
'--num_cpus',
type=int,
default=os.environ['SM_NUM_CPUS'],
)
self.parser.add_argument(
'--user_args',
default=json.loads(os.environ['SM_USER_ARGS']),
)
self.parser.add_argument(
'--input_dir',
type=str,
default=os.environ['SM_INPUT_DIR'],
)
self.parser.add_argument(
'--input_config_dir',
type=Path,
default=os.environ['SM_INPUT_CONFIG_DIR'],
)
self.parser.add_argument(
'--output_dir',
type=Path,
default=os.environ['SM_OUTPUT_DIR'],
)
# Extra arguments
self.run_tag = datetime.datetime \
.fromtimestamp(time.time()) \
.strftime('%Y-%m-%d-%H-%M-%S')
self.parser.add_argument(
'--run_tag',
default=self.run_tag,
type=str,
help=f"Run tag (default: 'datetime.fromtimestamp')",
)
self.parser.add_argument(
'--environment',
type=str,
default=self.environment,
)
self.args = self.parser.parse_args()
def get_arguments(self) -> Dict[str, Any]:
# Not in AWS Sagemaker arguments
self.parser.add_argument(
'--test_size',
default=0.2,
type=float,
help="Test dataset size (default: '0.2')",
)
self.parser.add_argument(
'--random_seed',
default=42,
type=int,
help="Random number for initialization (default: '42')",
)
self.parser.add_argument(
'--secrets',
type=YAMLFile.parse_string,
default='',
help="An yaml formated string (default: '')"
)
self.parser.add_argument(
'--bucket_name',
type=str,
default='',
help="Bucket name of a remote storage (default: '')"
)
self.args = self.parser.parse_args(self.args.user_args)
return self.args
和train
的entry_script会像这样开始:
#!/usr/bin/env python
from utils.arg_parser import TrainArgParser
if __name__ == '__main__':
logger.info(f"Begin train.py")
arg_parser = TrainArgParser()
args = arg_parser.get_arguments()
print(args)
这应该输出如下内容:
Namespace(bucket_name='', channel_names=['training'], environment='Docker', input_config_dir=PosixPath('/opt/ml/input/config'), input_dir='/opt/ml/input', model_dir='/opt/ml/model', num_cpus=8, num_gpus=1, output_dir=PosixPath('/opt/ml/output'), random_seed=42, run_tag='2020-03-11-22-18-21', secrets={}, test_size=0.2, user_args=['--test_size', '0.2', '--random_seed', '42'])
但是如果AWS Sagemaker将SM_USER_ARGS
和SM_HPS
视为同一事物,那就没用了。 :(