argtyped 0.3.0

Command line arguments, with types

argtyped: Command Line Argument Parser, with Types

argtyped is an command line argument parser with that relies on type annotations. It is built on argparse, the command line argument parser library built into Python. Compared with argparse, this library gives you:

• More concise and intuitive syntax, less boilerplate code.
• Type checking and IDE auto-completion for command line arguments.
• A drop-in replacement for argparse in most cases.

Installation

Install stable release from PyPI:

pip install argtyped

Or, install the latest commit from GitHub:

pip install git+https://github.com/huzecong/argtyped.git

Usage

With argtyped, you can define command line arguments in a syntax similar to typing.NamedTuple. The syntax is intuitive and can be illustrated with an example:

from typing import Optional
from typing_extensions import Literal  # or directly import from `typing` in Python 3.8+

from argtyped import Arguments, Switch
from argtyped import Enum, auto

class LoggingLevels(Enum):
    Debug = auto()
    Info = auto()
    Warning = auto()
    Error = auto()
    Critical = auto()

class MyArguments(Arguments):
    model_name: str         # required argument of `str` type
    hidden_size: int = 512  # `int` argument with default value of 512

    activation: Literal['relu', 'tanh', 'sigmoid'] = 'relu'  # argument with limited choices
    logging_level: LoggingLevels = LoggingLevels.Info        # using `Enum` class as choices

    use_dropout: Switch = True  # switch argument, enable with "--use-dropout" and disable with "--no-use-dropout"
    dropout_prob: Optional[float] = 0.5  # optional argument, "--dropout-prob=none" parses into `None`

args = MyArguments()

This is equivalent to the following code with Python built-in argparse:

import argparse
from enum import Enum

class LoggingLevels(Enum):
    Debug = "debug"
    Info = "info"
    Warning = "warning"
    Error = "error"
    Critical = "critical"

parser = argparse.ArgumentParser()

parser.add_argument("--model-name", type=str, required=True)
parser.add_argument("--hidden-size", type=int, default=512)

parser.add_argument("--activation", choices=["relu", "tanh", "sigmoid"], default="relu")
parser.add_argument("--logging-level", choices=list(LoggingLevels), type=LoggingLevels, default="info")

parser.add_argument("--use-dropout", action="store_true", dest="use_dropout", default=True)
parser.add_argument("--no-use-dropout", action="store_false", dest="use_dropout")
parser.add_argument("--dropout-prob", type=lambda s: None if s.lower() == 'none' else float(s), default=0.5)

args = parser.parse_args()

Save the code into a file named main.py. Suppose the following arguments are provided:

python main.py \
    --model-name LSTM \
    --activation sigmoid \
    --logging-level debug \
    --no-use-dropout \
    --dropout-prob none

Then the parsed arguments will be equivalent to the following structure returned by argparse:

argparse.Namespace(
    model_name="LSTM", hidden_size=512, activation="sigmoid", logging_level="debug",
    use_dropout=False, dropout_prob=None)

Arguments can also be pretty-printed:

>>> print(args)
<class '__main__.MyArguments'>
╔═════════════════╤══════════════════════════════════╗
║ Arguments       │ Values                           ║
╠═════════════════╪══════════════════════════════════╣
║ model_name      │ 'LSTM'                           ║
║ hidden_size     │ 512                              ║
║ activation      │ 'sigmoid'                        ║
║ logging_level   │ <MyLoggingLevels.Debug: 'debug'> ║
║ use_dropout     │ False                            ║
║ dropout_prob    │ None                             ║
║ label_smoothing │ 0.1                              ║
║ some_true_arg   │ True                             ║
║ some_false_arg  │ False                            ║
╚═════════════════╧══════════════════════════════════╝

It is recommended though to use the args.to_string() method, which gives you control of the table width.

Reference

The argtyped.Arguments Class

The argtyped.Arguments class is main class of the package, from which you should derive your custom class that holds arguments. Each argument takes the form of a class attribute, with its type annotation and an optional default value.

When an instance of your custom class is initialized, the command line arguments are parsed from sys.argv into values with your annotated types. You can also provide the list of strings to parse by passing them as the parameter.

The parsed arguments are stored in an object of your custom type. This gives you arguments that can be auto-completed by the IDE, and type-checked by a static type checker like mypy.

The following example illustrates the keypoints:

class MyArgs(argtyped.Arguments):
    # name: type [= default_val]
    value: int = 0

args = MyArgs()                    # equivalent to `parser.parse_args()`
args = MyArgs(["--value", "123"])  # equivalent to `parser.parse_args(["--value", "123"])
assert isinstance(args, MyArgs)

Arguments.to_dict(self)

Convert the set of arguments to a dictionary (OrderedDict).

Arguments.to_string(self, width: Optional[int] = None, max_width: Optional[int] = None)

Represent the arguments as a table.

• width: Width of the printed table. Defaults to None, which fits the table to its contents. An exception is raised when the table cannot be drawn with the given width.
• max_width: Maximum width of the printed table. Defaults to None, meaning no limits. Must be None if width is not None.

argtyped.argument_specs

Return a dictionary mapping argument names to their specifications, represented as the argtyped.ArgumentSpec type. This is useful for programmatically accessing the list of arguments.

Argument Types

To summarize, whatever works for argparse works here. The following types are supported:

• Built-in types such as int, float, str.

• Boolean type bool. Accepted values (case-insensitive) for True are: y, yes, true, ok; accepted values for False are: n, no, false.

• Choice types Literal[...] or Choices[...]. A choice argument is essentially an str argument with limited choice of values. The ellipses can be filled with a tuple of strs, or an expression that evaluates to a list of strs:

  from argtyped import Arguments, Choices
  from typing import List
  from typing_extensions import Literal
  
  def logging_levels() -> List[str]:
      return ["debug", "info", "warning", "error"]
  
  class MyArgs(Arguments):
      foo: Literal["debug", "info", "warning", "error"]  # 4 choices
      bar: Choices[logging_levels()]                     # the same 4 choices
  
  # argv: ["--foo=debug", "--bar=info"] => foo="debug", bar="info"
  

  This is equivalent to the choices keyword in argparse.add_argument.

  Note: The choice type was previously named Choices. This is deprecated in favor of the Literal type introduced in Python 3.8 and back-ported to 3.6 and 3.7 in the typing_extensions library. Please see Notes for a discussing on the differences between the two.

• Enum types derived from enum.Enum. It is recommended to use argtyped.Enum which uses the instance names as values:

  from argtyped import Enum
  
  class MyEnum(Enum):
      Debug = auto()    # "debug"
      Info = auto()     # "info"
      Warning = auto()  # "warning"
  

• Switch types Switch. Switch arguments are like bool arguments, but they don't take values. Instead, a switch argument switch requires --switch to enable and --no-switch to disable:

  from argtyped import Arguments, Switch
  
  class MyArgs(Arguments):
      switch: Switch = True
      bool_arg: bool = False
  
  # argv: []                                 => flag=True,  bool_arg=False
  # argv: ["--switch", "--bool-arg=false"]   => flag=True,  bool_arg=False
  # argv: ["--no-switch", "--bool-arg=true"] => flag=False, bool_arg=True
  # argv: ["--switch=false"]                 => WRONG
  # argv: ["--no-bool-arg"]                  => WRONG
  

• List types List[T], where T is any supported type except switch types. List arguments allow passing multiple values on the command line following the argument flag, it is equivalent to setting nargs="*" in argparse.

  Although there is no built-in support for other nargs settings such as "+" (one or more) or N (fixed number), you can add custom validation logic by overriding the __init__ method in your Arguments subclass.

• Optional types Optional[T], where T is any supported type except list or switch types. An optional argument will be filled with None if no value is provided. It could also be explicitly set to None by using none as value in the command line:

  from argtyped import Arguments
  from typing import Optional
  
  class MyArgs(Arguments):
      opt_arg: Optional[int]  # implicitly defaults to `None`
  
  # argv: []                 => opt_arg=None
  # argv: ["--opt-arg=1"]    => opt_arg=1
  # argv: ["--opt-arg=none"] => opt_arg=None
  

• Any other type that takes a single str as __init__ parameters. It is also theoretically possible to use a function that takes an str as input, but it's not recommended as it's not type-safe.

Notes

• Advanced argparse features such as subparsers, groups, argument lists, and custom actions are not supported.
• Using switch arguments may result in name clashes: if a switch argument has name arg, there can be no argument with the name no_arg.
• Optional types:
  • Optional can be used with Literal, but cannot be used with Choices:

    from argtyped import Arguments, Choices
    from typing import Literal, Optional
    
    class MyArgs(Arguments):
        foo: Optional[Literal["a", "b"]]  # valid
        bar: Optional[Choices["a", "b"]]  # invalid
        baz: Choices["a", "b", "none"]    # not elegant but it works
    

  • Optional[str] would parse a value of "none" (case-insensitive) into None.
• List types:
  • List[Optional[T]] is a valid type. For example:

    from argtyped import Arguments
    from typing import List, Literal, Optional
    
    class MyArgs(Arguments):
        foo: List[Optional[Literal["a", "b"]]] = ["a", None, "b"]  # valid
    
    # argv: ["--foo", "a", "b", "none", "a", "b"] => foo=["a", "b", None, "a", "b"]
    

  • List types cannot be nested inside a list or an optional type. Types such as Optional[List[int]] and List[List[int]] are not accepted.
• Choices vs Literal:
  • Choices is deprecated. In general, you should prefer Literal to Choices.
  • When all choices are defined as string literals, the two types are interchangeable.
  • Pros for Choices:
    • The Choices parameter can be an expression that evaluate to an iterable of strings, while Literal only supports string literals as parameters.
    • Literal requires installing the typing_extensions package in Python versions prior to 3.8. This package is not listed as a prerequisite of argtyped so you must manually install it.
  • Pros for Literal:
    • Literal is a built-in type supported by type-checkers and IDEs. You can get better type inference with Literal, and the IDE won't warn you that your choices are "undefined" (because it interprets the string literals as forward references).
    • Literal can be combined with Optional.

Under the Hood

This is what happens under the hood:

1. When a subclass of argtyped.Arguments is constructed, type annotations and class-level attributes (i.e., the default values) are collected to form argument declarations.
2. After verifying the validity of declared arguments, argtyped.ArgumentSpec are created for each argument and stored within the subclass as the __arguments__ class attribute.
3. When an instance of the subclass is initialized, if it's the first time, an instance of argparse.ArgumentParser is created and arguments are registered with the parser. The parser is cached in the subclass as the __parser__ attribute.
4. The parser's parse_args method is invoked with either sys.argv or strings provided as parameters, returning parsed arguments.
5. The parsed arguments are assigned to self (the instance of the Arguments subclass being initialized).

Todo

• Support action="append" or action="extend" for List[T] types.
  • Technically this is not a problem, but there's no elegant way to configure whether this behavior is desired.
• Throw (suppressable) warnings on using non-type callables as types.
• Support converting an attrs class into Arguments.
• Support forward references in type annotations.