Skip to content

Latest commit

 

History

History
2613 lines (1854 loc) · 89.6 KB

DOCUMENTATION.rst

File metadata and controls

2613 lines (1854 loc) · 89.6 KB
Raw

Basic usage

There are multiple ways of using jsonargparse. One is to construct low level parsers (see parsers) being almost a drop in replacement of argparse. However, argparse is too verbose and leads to unnecessary duplication. The simplest and recommended way of using jsonargparse is by using the .CLI function, which has the benefit of minimizing boilerplate code. A simple example is:

from jsonargparse import CLI

def command(name: str, prize: int = 100):

"""Prints the prize won by a person.

Args:

name: Name of winner. prize: Amount won.

""" print(f"{name} won {prize}€!")

if __name__ == "__main__":

CLI(command)

Note that the name and prize parameters have type hints and are described in the docstring. These are shown in the help of the command line tool. In a shell you could see the help and run a command as follows:

$ python example.py --help
...
Prints the prize won by a person:
  name                  Name of winner. (required, type: str)
  --prize PRIZE         Amount won. (type: int, default: 100)

$ python example.py Lucky --prize=1000
Lucky won 1000€!

Note

Parsing of docstrings is an optional feature. For this example to work as shown, jsonargparse needs to be installed with the signatures extras require as explained in section installation.

When .CLI receives a single class, the first arguments are for parameters to instantiate the class, then a method name is expected (i.e. methods become sub-commands) and the remaining arguments are for parameters of this method. An example would be:

from random import randint from jsonargparse import CLI

class Main:
def __init__(self, max_prize: int = 100):

""" Args: max_prize: Maximum prize that can be awarded. """ self.max_prize = max_prize

def person(self, name: str):

""" Args: name: Name of winner. """ return f"{name} won {randint(0, self.max_prize)}€!"

if __name__ == "__main__":

print(CLI(Main))

Then in a shell you could run:

$ python example.py --max_prize=1000 person Lucky
Lucky won 632€!
hide

>>> CLI(Main, args=["--max_prize=1000", "person", "Lucky"]) # doctest: +ELLIPSIS 'Lucky won ...€!'

If the class given does not have any methods, there will be no sub-commands and .CLI will return an instance of the class. For example:

from dataclasses import dataclass from jsonargparse import CLI

@dataclass class Settings: name: str prize: int = 100

if __name__ == "__main__":

print(CLI(Settings, as_positional=False))

Then in a shell you could run:

$ python example.py --name=Lucky
Settings(name='Lucky', prize=100)
hide

>>> CLI(Settings, as_positional=False, args=["--name=Lucky"]) # doctest: +ELLIPSIS Settings(name='Lucky', prize=100)

Note the use of as_positional=False to make required arguments as non-positional.

If more than one function is given to .CLI, then any of them can be run via sub-commands similar to the single class example above, i.e. example.py function [arguments] where function is the name of the function to execute. If multiple classes or a mixture of functions and classes is given to .CLI, to execute a method of a class, two levels of sub-commands are required. The first sub-command would be the name of the class and the second the name of the method, i.e. example.py class [init_arguments] method [arguments].

Arbitrary levels of sub-commands with custom names can be defined by providing a dict. For example:

class Raffle:
def __init__(self, prize: int):

self.prize = prize

def __call__(self, name: str):

return f"{name} won {self.prize}€!"

components = {
"weekday": {

"_help": "Raffles for weekdays", "tier1": Raffle(prize=100), "tier2": Raffle(prize=50),

}, "weekend": { "_help": "Raffles for weekends", "tier1": Raffle(prize=300), "tier2": Raffle(prize=75), },

}

if __name__ == "__main__":

print(CLI(components))

Then in a shell:

$ python example.py weekend tier1 Lucky
Lucky won 300€!
hide

>>> CLI(components, args=["weekend", "tier1", "Lucky"]) 'Lucky won 300€!'

Note

The examples above are extremely simple, only defining parameters with str and int type hints. The true power of jsonargparse is its support for a wide range of types, see type-hints. It is even possible to use general classes as type hints, allowing to easily implement configurable dependency injection (object composition), see sub-classes.

Writing configuration files

All tools implemented with the .CLI function have the --config option to provide settings in a config file (more details in configuration-files). This becomes very useful when the number of configurable parameters is large. To ease the writing of config files, there is also the option --print_config which prints to standard output a yaml with all settings that the tool supports with their default values. Users of the tool can be advised to follow the following steps:

# Dump default configuration to have as reference
python example.py --print_config > config.yaml
# Modify the config as needed (all default settings can be removed)
nano config.yaml
# Run the tool using the adapted config
python example.py --config config.yaml

Comparison to Fire

The .CLI feature is similar to and inspired by Fire. However, there are fundamental differences. First, the purpose is not allowing to call any python object from the command line. It is only intended for running functions and classes specifically written for this purpose. Second, the arguments are expected to have type hints, and the given values will be validated according to these. Third, the return values of the functions are not automatically printed. .CLI returns the value and it is up to the developer to decide what to do with it.

Tutorials

Parsers

An argument parser is created just like it is done with python's argparse. You import the module, create a parser object and then add arguments to it. A simple example would be:

from jsonargparse import ArgumentParser

parser = ArgumentParser(prog="app", description="Description for my app.") parser.add_argument("--opt1", type=int, default=0, help="Help for option 1.") parser.add_argument("--opt2", type=float, default=1.0, help="Help for option 2.")

After creating the parser, you can use it to parse command line arguments with the :py.ArgumentParser.parse_args function, after which you get an object with the parsed values or defaults available as attributes. For illustrative purposes giving to parse_args a list of arguments (instead of automatically getting them from the command line arguments), with the parser shown above you would observe:

>>> cfg = parser.parse_args(["--opt2", "2.3"]) >>> cfg.opt1, type(cfg.opt1) (0, <class 'int'>) >>> cfg.opt2, type(cfg.opt2) (2.3, <class 'float'>)

If the parsing fails the standard behavior is that the usage is printed and the program is terminated. Alternatively you can initialize the parser with exit_on_error=False in which case an .ArgumentError is raised.

Override order

Final parsed values depend on different sources, namely: source code, command line arguments, configuration-files and environment-variables. Values are overridden based on the following precedence:

  1. Defaults defined in the source code.
  2. Existing default config files in the order defined in default_config_files, e.g. ~/.config/myapp.yaml.
  3. Full config environment variable, e.g. APP_CONFIG.
  4. Individual key environment variables, e.g. APP_OPT1.
  5. Command line arguments in order left to right (might include config files).

Depending on the parse method used (see .ArgumentParser) and how the parser was built, some of the options above might not apply. Parsing of environment variables must be explicitly enabled, except if using :py.ArgumentParser.parse_env. If the parser does not have an action="config" argument, then there is no parsing of a full config environment variable or a way to provide a config file from command line.

Capturing parsers

It can be common practice to have a function that implements an entire CLI or a function that constructs a parser conditionally based on some parameters and then parses. For example, one might have:

from jsonargparse import ArgumentParser

def main_cli():

parser = ArgumentParser() ... cfg = parser.parse_args() ...

if __name__ == "__main__":

main_cli()

For some use cases it is necessary to get an instance of the parser object, without doing any parsing. For instance sphinx-argparse can be used to include the help of CLIs in automatically generated documentation of a package. To use sphinx-argparse it is necessary to have a function that returns the parser. Having a CLI function this could be easily implemented with .capture_parser as follows:

from jsonargparse import capture_parser

def get_parser():

return capture_parser(main_cli)

Note

The official way to obtain the parser for command line tools based on .CLI is by using .capture_parser.

Functions as type

Using a function as a type, like int_or_off below, is supported though discouraged. A basic requirement is that the function be idempotent, i.e., applying the function two or more times should not modify the value. Instead of a function, it is recommended to implement a type, see custom-types.

# either int larger than zero or 'off' string def int_or_off(x): return x if x == "off" else int(x)

parser.add_argument("--int_or_off", type=int_or_off)

Type hints

An important feature of jsonargparse is a wide support the argument types and respective validation. This extended support makes use of Python's type hint syntax. For example, an argument that can be None or a float in the range (0, 1) or a positive int could be added using a type hint as follows:

from typing import Optional, Union from jsonargparse.typing import PositiveInt, OpenUnitInterval

parser.add_argument("--op", type=Optional[Union[PositiveInt, OpenUnitInterval]])

The types in :pyjsonargparse.typing are included for convenience since they are useful in argument parsing use cases and not available in standard python. However, there is no need to use jsonargparse specific types.

A wide range of type hints are supported and with arbitrary complexity/nesting. Some notes about this support are:

  • Nested types are supported as long as at least one child type is supported. By nesting it is meant child types inside List, Dict, etc. There is no limit in nesting depth.
  • Postponed evaluation of types PEP 563 (i.e. from __future__ import annotations) is supported. Also supported on python<=3.9 are PEP 585 (i.e. list[<type>], dict[<type>], ... instead of List[<type>], Dict[<type>], ...) and 604 (i.e. <type> | <type> instead of Union[<type>, <type>]).
  • Types that use components imported inside TYPE_CHECKING blocks are supported.
  • Resolving of forward references in types is supported.
  • Fully supported types are: str, bool (more details in boolean-arguments), int, float, Decimal, complex, bytes/bytearray (Base64 encoding), range, List (more details in list-append), Iterable, Sequence, Any, Union, Optional, Type, Enum, PathLike, UUID, timedelta, restricted types as explained in sections restricted-numbers and restricted-strings and paths and URLs as explained in sections parsing-paths and parsing-urls.
  • Dict, Mapping, MutableMapping, and TypedDict are supported but only with str or int keys. For more details see dict-items.
  • Tuple, Set and MutableSet are supported even though they can't be represented in json distinguishable from a list. Each Tuple element position can have its own type and will be validated as such. Tuple with ellipsis (Tuple[type, ...]) is also supported. In command line arguments, config files and environment variables, tuples and sets are represented as an array.
  • To set a value to None it is required to use null since this is how json/yaml defines it. To avoid confusion in the help, NoneType is displayed as null. For example a function argument with type and default Optional[str] = None would be shown in the help as type: Union[str, null], default: null.
  • Normal classes can be used as a type, which are specified with a dict containing class_path and optionally init_args. :py.ArgumentParser.instantiate_classes can be used to instantiate all classes in a config object. For more details see sub-classes.
  • dataclasses are supported even when nested. Final classes, attrs' define decorator, and pydantic's dataclass decorator and BaseModel classes are supported and behave like standard dataclasses. For more details see dataclass-like. If a dataclass is mixed inheriting from a normal class, it is considered a subclass type instead of a dataclass.
  • User-defined Generic types are supported. For more details see generic-types.
  • Annotated types are supported. If the metadata corresponds to a pydantic type, this is used for validation.
  • pydantic.SecretStr type is supported with the expected behavior of not serializing the actual value. There is also jsonargparse.typing.SecretStr to support the same behavior without the need of a dependency.
  • Callable is supported by either giving a dot import path to a callable object or by giving a dict with a class_path and optionally init_args entries. The specified class must either instantiate into a callable or be a subclass of the return type of the callable. For these cases running :py.ArgumentParser.instantiate_classes will instantiate the class or provide a function that returns the instance of the class. For more details see callable-type. Currently the callable's argument and return types are not validated.
  • TypeAliasType is supported with values parsed as the aliased type and the alias shown as the argument type in help.

Restricted numbers

It is quite common that when parsing a number, its range should be limited. To ease these cases the module jsonargparse.typing includes some predefined types and a function .restricted_number_type to define new types. The predefined types are: .PositiveInt, .NonNegativeInt, .PositiveFloat, .NonNegativeFloat, .ClosedUnitInterval and .OpenUnitInterval. Examples of usage are:

from jsonargparse.typing import PositiveInt, PositiveFloat, restricted_number_type

# float larger than zero parser.add_argument("--op1", type=PositiveFloat) # between 0 and 10 from_0_to_10 = restricted_number_type("from_0_to_10", int, [(">=", 0), ("<=", 10)]) parser.add_argument("--op2", type=from_0_to_10)

Restricted strings

Similar to the restricted numbers, there is a function to create string types that are restricted to match a given regular expression: .restricted_string_type. A predefined type is .Email which is restricted so that it follows the normal email pattern. For example to add an argument required to be exactly four uppercase letters:

from jsonargparse.typing import Email, restricted_string_type

CodeType = restricted_string_type("CodeType", "^[A-Z]{4}$") parser.add_argument("--code", type=CodeType) parser.add_argument("--email", type=Email)

Parsing paths

For some use cases it is necessary to parse file paths, checking its existence and access permissions, but not necessarily opening the file. Moreover, a file path could be included in a config file as relative with respect to the config file's location. After parsing it should be easy to access the parsed file path without having to consider the location of the config file. To help in these situations jsonargparse includes a type generator .path_type, some predefined types (e.g. .Path_fr).

For example suppose you have a directory with a configuration file app/config.yaml and some data app/data/info.db. The contents of the yaml file is the following:

# File: config.yaml
databases:
  info: data/info.db

To create a parser that checks that the value of databases.info is a file that exists and is readable, the following could be done:

paths

cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) os.mkdir("app") os.mkdir("app/data") with open("app/config.yaml", "w") as f: f.write("databases:n info: data/info.dbn") with open("app/data/info.db", "w") as f: f.write("infon")

paths

os.chdir(cwd) shutil.rmtree(tmpdir)

paths

from jsonargparse import ArgumentParser from jsonargparse.typing import Path_fr

parser = ArgumentParser() parser.add_argument("--databases.info", type=Path_fr) cfg = parser.parse_path("app/config.yaml")

The fr in the type are flags that stand for file and readable. After parsing, the value of databases.info will be an instance of the .Path_fr class that allows to get both the original relative path as included in the yaml file, or the corresponding absolute path:

paths

>>> cfg.databases.info.relative 'data/info.db' >>> cfg.databases.info.absolute # doctest: +ELLIPSIS '/.../app/data/info.db'

Likewise directories can be parsed using the .Path_dw type, which would require a directory to exist and be writeable. New path types can be created using the .path_type function. For example to create a type for files that must exist and be both readable and writeable, the command would be Path_frw = path_type('frw'). If the file app/config.yaml is not writeable, then using the type to cast Path_frw('app/config.yaml') would raise a TypeError: File is not writeable exception. For more information of all the mode flags supported, refer to the documentation of the .Path class.

Types created with .path_type have as base class .Path. This class implements the os.PathLike protocol, using the absolute version as the actual path, thus for the previous example:

paths

>>> os.fspath(cfg.databases.info) # doctest: +ELLIPSIS '/.../app/data/info.db'

The content of a file that a .Path instance references can be read by using the :py.Path.get_content method. For the previous example would be info_db = cfg.databases.info.get_content().

An argument with a path type can be given nargs='+' to parse multiple paths. But it might also be wanted to parse a list of paths found in a plain text file or from stdin. For this add the argument with type List[<path_type>] and enable_path=True. To read from stdin give the special string '-'. Example:

path_list

cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) pathlib.Path("paths.lst").write_text("paths.lstn")

parser = ArgumentParser()

stdin = sys.stdin sys.stdin = StringIO("paths.lstn")

path_list

sys.stdin = stdin os.chdir(cwd) shutil.rmtree(tmpdir)

path_list

from jsonargparse.typing import Path_fr

parser.add_argument("--list", type=List[Path_fr], enable_path=True) cfg = parser.parse_args(["--list", "paths.lst"]) # File with list of paths cfg = parser.parse_args(["--list", "-"]) # List of paths from stdin

If nargs='+' is given to add_argument with List[<path_type>] and enable_path=True then for each argument a list of paths is generated.

Note

No all features of the .Path class are supported in windows.

Parsing URLs

The .path_type function also supports URLs which after parsing, the :py.Path.get_content method can be used to perform a GET request to the corresponding URL and retrieve its content. For this to work the requests python package is required. Alternatively, .path_type can also be used for fsspec supported file systems. The respective optional package(s) will be installed along with jsonargparse if installed with the urls or fsspec extras require as explained in section installation.

The 'u' flag is used to parse URLs using requests and the flag 's' to parse fsspec file systems. For example if it is desired that an argument can be either a readable file or URL, the type would be created as Path_fur = path_type('fur'). If the value appears to be a URL, a HEAD request would be triggered to check if it is accessible. To get the content of the parsed path, without needing to care if it is a local file or a URL, the :py.Path.get_content method Scan be used.

If you import from jsonargparse import set_config_read_mode and then run set_config_read_mode(urls_enabled=True) or set_config_read_mode(fsspec_enabled=True), the following functions and classes will also support loading from URLs: :py.ArgumentParser.parse_path, :py.ArgumentParser.get_defaults (default_config_files argument), action="config", .ActionJsonSchema, .ActionJsonnet and .ActionParser. This means that a tool that can receive a configuration file via action="config" is able to get the content from a URL, thus something like the following would work:

my_tool.py --config http://example.com/config.yaml

Note

Relative paths inside a remote path are parsed as remote. For example, for a relative path model/state_dict.pt found inside s3://bucket/config.yaml, its parsed absolute path becomes s3://bucket/model/state_dict.pt.

Booleans

Parsing boolean arguments is very common, however, the original argparse only has a limited support for them, via store_true and store_false. Furthermore unexperienced users might mistakenly use type=bool which would not provide the intended behavior.

With jsonargparse adding an argument with type=bool the intended action is implemented. If given as values {'yes', 'true'} or {'no', 'false'} the corresponding parsed values would be True or False. For example:

boolean

parser = ArgumentParser()

boolean

>>> parser.add_argument("--op1", type=bool, default=False) # doctest: +IGNORE_RESULT >>> parser.add_argument("--op2", type=bool, default=True) # doctest: +IGNORE_RESULT >>> parser.parse_args(["--op1", "yes", "--op2", "false"]) Namespace(op1=True, op2=False)

Sometimes it is also useful to define two paired options, one to set True and the other to set False. The .ActionYesNo class makes this straightforward. A couple of examples would be:

yes_no

parser = ArgumentParser()

yes_no

from jsonargparse import ActionYesNo

# --opt1 for true and --no_opt1 for false. parser.add_argument("--op1", action=ActionYesNo) # --with-opt2 for true and --without-opt2 for false. parser.add_argument("--with-op2", action=ActionYesNo(yes_prefix="with-", no_prefix="without-"))

If the .ActionYesNo class is used in conjunction with nargs='?' the options can also be set by giving as value any of {'true', 'yes', 'false', 'no'}.

Enum arguments

Another case of restricted values is string choices. In addition to the common choices given as a list of strings, it is also possible to provide as type an Enum class. This has the added benefit that strings are mapped to some desired values. For example:

enum

parser = ArgumentParser()

enum

>>> import enum >>> class MyEnum(enum.Enum): ... choice1 = -1 ... choice2 = 0 ... choice3 = 1 ... >>> parser.add_argument("--op", type=MyEnum) # doctest: +IGNORE_RESULT >>> parser.parse_args(["--op=choice1"]) Namespace(op=<MyEnum.choice1: -1>)

List append

As detailed before, arguments with List type are supported. By default when specifying an argument value, the previous value is replaced, and this also holds for lists. Thus, a parse such as parser.parse_args(['--list=[1]', '--list=[2, 3]']) would result in a final value of [2, 3]. However, in some cases it might be decided to append to the list instead of replacing. This can be achieved by adding + as suffix to the argument key, for example:

append

parser = ArgumentParser()

class MyBaseClass:

pass

append

>>> parser.add_argument("--list", type=List[int]) # doctest: +IGNORE_RESULT >>> parser.parse_args(["--list=[1]", "--list+=[2, 3]"]) Namespace(list=[1, 2, 3]) >>> parser.parse_args(["--list=[4]", "--list+=5"]) Namespace(list=[4, 5])

Append is also supported in config files. For instance the following two config files would first assign a list and then append to this list:

# config1.yaml
list:
- 1
# config2.yaml
list+:
- 2
- 3

Appending works for any type for the list elements. Lists with class type elements (see sub-classes) are also supported. To append to the list, first append a new class by using the + suffix. Then init_args for this class are specified like if the type wasn't a list, since the arguments are applied to the last class in the list. Take for example that an argument is added to a parser as:

append

parser.add_argument("--list_of_instances", type=List[MyBaseClass])

Thanks to the short notation, command line arguments don't require to specify class_path and init_args. Thus, multiple classes can be appended and its arguments set as follows:

python tool.py \
  --list_of_instances+={CLASS_1_PATH} \
  --list_of_instances.{CLASS_1_ARG_1}=... \
  --list_of_instances.{CLASS_1_ARG_2}=... \
  --list_of_instances+={CLASS_2_PATH} \
  --list_of_instances.{CLASS_2_ARG_1}=... \
  ...
  --list_of_instances+={CLASS_N_PATH} \
  --list_of_instances.{CLASS_N_ARG_1}=... \
  ...

Once a new class has been appended to the list, it is not possible to modify the arguments of a previous class. This limitation is by intention since it forces classes and its arguments to be defined in order, making the command line call intuitive to write and understand.

Dict items

When an argument has Dict as type, the value can be set using json format, e.g.:

dict_items

parser = ArgumentParser()

dict_items

>>> parser.add_argument("--dict", type=dict) # doctest: +IGNORE_RESULT >>> parser.parse_args(['--dict={"key1": "val1", "key2": "val2"}']) Namespace(dict={'key1': 'val1', 'key2': 'val2'})

Similar to lists, providing a second argument with value a json dict completely replaces the previous value. Setting individual dict items without replacing can be achieved as follows:

dict_items

>>> parser.parse_args(["--dict.key1=val1", "--dict.key2=val2"]) Namespace(dict={'key1': 'val1', 'key2': 'val2'})

Dataclass-like classes

In contrast to subclasses, which requires the user to provide a class_path, in some cases it is not expected to have subclasses. In this case the init args are given directly in a dictionary without specifying a class_path. This is the behavior for standard dataclasses, final classes, attrs' define decorator, and pydantic's dataclass decorator and BaseModel classes.

As an example, take a class that is decorated with .final, meaning that it shouldn't be subclassed. The code below would accept the corresponding yaml structure.

final_classes

cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) with open("config.yaml", "w") as f: f.write("data:\n number: 8n accepted: truen")

final_classes

os.chdir(cwd) shutil.rmtree(tmpdir)

final_classes

from jsonargparse.typing import final

@final class FinalClass: def __init__(self, number: int = 0, accepted: bool = False): ...

parser = ArgumentParser() parser.add_argument("--data", type=FinalClass) cfg = parser.parse_path("config.yaml")

data:
  number: 8
  accepted: true

Generic types

Classes that inherit from typing.Generic, also known as user-defined generic types, are supported. Take for example a point in 2D:

generic_types

parser = ArgumentParser()

generic_types

from typing import Generic, TypeVar

Number = TypeVar("Number", float, complex)

@dataclass class Point2d(Generic[Number]): x: Number = 0.0 y: Number = 0.0

Parsing complex-valued points would be:

generic_types

>>> parser.add_argument("--point", type=Point2d[complex]) # doctest: +IGNORE_RESULT >>> parser.parse_args(["--point.x=(1+2j)"]).point Namespace(x=(1+2j), y=0.0)

Callable type

When using Callable as type, the parser accepts several options. The first option is the import path of a callable object, for example:

callable

parser = ArgumentParser()

callable

parser.add_argument("--callable", type=Callable) parser.parse_args(["--callable=time.sleep"])

A second option is a class that once instantiated becomes callable:

callable

class OffsetSum:
def __init__(self, offset: int):

self.offset = offset

def __call__(self, value: int):

return self.offset + value

callable

doctest_mock_class_in_main(OffsetSum)

callable

>>> value = { ... "class_path": "__main__.OffsetSum", ... "init_args": { ... "offset": 3, ... }, ... }

>>> cfg = parser.parse_args(["--callable", str(value)]) >>> cfg.callable Namespace(class_path='__main__.OffsetSum', init_args=Namespace(offset=3)) >>> init = parser.instantiate_classes(cfg) >>> init.callable(5) 8

The third option is only applicable when the type is a callable that has a class as return type or a Union including a class. This is useful to support dependency injection for classes that require a parameter that is only available after injection. The parser supports this automatically by providing a function that receives this parameter and returns the instance of the class. Take for example the classes:

callable

class Optimizer:
def __init__(self, params: Iterable):

self.params = params

class SGD(Optimizer):
def __init__(self, params: Iterable, lr: float):

super().__init__(params) self.lr = lr

callable

doctest_mock_class_in_main(SGD)

A possible parser and callable behavior would be:

callable

>>> value = { ... "class_path": "SGD", ... "init_args": { ... "lr": 0.01, ... }, ... }

>>> parser.add_argument("--optimizer", type=Callable[[Iterable], Optimizer]) # doctest: +IGNORE_RESULT >>> cfg = parser.parse_args(["--optimizer", str(value)]) >>> cfg.optimizer Namespace(class_path='__main__.SGD', init_args=Namespace(lr=0.01)) >>> init = parser.instantiate_classes(cfg) >>> optimizer = init.optimizer([1, 2, 3]) >>> isinstance(optimizer, SGD) True >>> optimizer.params, optimizer.lr ([1, 2, 3], 0.01)

Multiple arguments available after injection are also supported and can be specified the same way with a Callable type hint. For example, for two Iterable arguments, you can use the following syntax: Callable[[Iterable, Iterable], Type]. Please be aware that the arguments are passed as positional arguments, this means that the injected function would be called like function(value1, value2). Similarly, for a callable that accepts zero arguments, the syntax would be Callable[[], Type].

Note

When the Callable has a class return type, it is possible to specify the class_path giving only its name if imported before parsing, as explained in sub-classes-command-line.

If the same type above is used as type hint of a parameter of another class, a default can be set using a lambda, for example:

callable

class Model:
def __init__(

self, optimizer: Callable[[Iterable], Optimizer] = lambda p: SGD(p, lr=0.05),

):

self.optimizer = optimizer

Then a parser and behavior could be:

>>> parser.add_class_arguments(Model, 'model')
>>> cfg = parser.get_defaults()
>>> cfg.model.optimizer
Namespace(class_path='__main__.SGD', init_args=Namespace(lr=0.05))
>>> init = parser.instantiate_classes(cfg)
>>> optimizer = init.model.optimizer([1, 2, 3])
>>> optimizer.params, optimizer.lr
([1, 2, 3], 0.05)

See ast-resolver for limitations of lambda defaults in signatures. Providing a lambda default to :py.ArgumentParser.add_argument does not work since there is no AST resolving. In this case, a dict with class_path and init_args can be used as default.

Registering types

With the .register_type function it is possible to register additional types for use in jsonargparse parsers. If the type class can be instantiated with a string representation and casting the instance to str gives back the string representation, then only the type class is given to .register_type. For example in the jsonargparse.typing package this is how complex numbers are registered: register_type(complex). For other type classes that don't have these properties, to register it might be necessary to provide a serializer and/or deserializer function. Including the serializer and deserializer functions, the registration of the complex numbers example is equivalent to register_type(complex, serializer=str, deserializer=complex).

A more useful example could be registering the datetime class. This case requires to give both a serializer and a deserializer as seen below.

from datetime import datetime from jsonargparse import ArgumentParser from jsonargparse.typing import register_type

def serializer(v):

return v.isoformat()

def deserializer(v):

return datetime.strptime(v, "%Y-%m-%dT%H:%M:%S")

register_type(datetime, serializer, deserializer)

parser = ArgumentParser() parser.add_argument("--datetime", type=datetime) parser.parse_args(["--datetime=2008-09-03T20:56:35"])

Note

The registering of types is only intended for simple types. By default any class used as a type hint is considered a sub-class (see sub-classes) which might be good for many use cases. If a class is registered with .register_type then the sub-class option is no longer available.

Creating custom types

It is possible to create new types and use them for parsing. Even though types can be created for specific CLI behaviors, it is recommended to create them such that they make sense independent of parsing. This is so that they can be used as type hints in functions and classes in order to improve the code in a more general sense. An alternative to creating types, can be to use pydantic types.

There are a few ways for creating types, the most simple being to implement a class. When creating a type, take as reference how basic types work, e.g. int. Properties of basic types are:

  • Casting a string creates an instance of the type, if the value is valid, e.g. int("1").
  • Casting a string raises a ValueError, if the value is not valid, e.g. int("a").
  • Casting an instance of the type to string gives back the string representation of the value, e.g. str(1) == "1".
  • Types are idempotent, i.e. casting an instance of the type to the type gives back the same value, e.g. int(1) == int(int(1)).

Once a type is created, it can be registered with .register_type. If the type follows the properties above, then there is no need to provide more parameters, just do register_type(MyType).

The .extend_base_type function can be useful for creating and registering new types in a single call. For example, creating a type for even integers could be done as:

from jsonargparse.typing import extend_base_type

def is_even(type_class, value):
if int(value) % 2 != 0:

raise ValueError(f"{value} is not even")

EvenInt = extend_base_type("EvenInt", int, is_even)

Then this type can be used in a parser as:

>>> parser = ArgumentParser() >>> parser.add_argument("--even_int", type=EvenInt) # doctest: +IGNORE_RESULT >>> parser.parse_args(["--even_int=2"]) Namespace(even_int=2)

When using custom types as a type hint, defaults must be casted so that static type checkers don't complain. For example:

def fn(value: EvenInt = EvenInt(2)):

...

Nested namespaces

A difference with respect to basic argparse is, that by using dot notation in the argument names, you can define a hierarchy of nested namespaces. For example you could do the following:

>>> parser = ArgumentParser(prog="app") >>> parser.add_argument("--lev1.opt1", default="from default 1") # doctest: +IGNORE_RESULT >>> parser.add_argument("--lev1.opt2", default="from default 2") # doctest: +IGNORE_RESULT >>> cfg = parser.get_defaults() >>> cfg.lev1.opt1 'from default 1' >>> cfg.lev1.opt2 'from default 2'

A group of nested options can be created by using a dataclass. This has the advantage that the same options can be reused in multiple places of a project. An example analogous to the one above would be:

from dataclasses import dataclass

@dataclass class Level1Options: """Level 1 options Args: opt1: Option 1 opt2: Option 2 """

opt1: str = "from default 1" opt2: str = "from default 2"

parser = ArgumentParser() parser.add_argument("--lev1", type=Level1Options, default=Level1Options())

The .Namespace class is an extension of the one from argparse, having some additional features. In particular, keys can be accessed like a dictionary either with individual keys, e.g. cfg['lev1']['opt1'], or a single one, e.g. cfg['lev1.opt1']. Also the class has a method :py.Namespace.as_dict that can be used to represent the nested namespace as a nested dictionary. This is useful for example for class instantiation.

Configuration files

An important feature of jsonargparse is the parsing of yaml/json files. The dot notation hierarchy of the arguments (see nested-namespaces) are used for the expected structure in the config files.

The :py.ArgumentParser.default_config_files property can be set when creating a parser to specify patterns to search for configuration files. For example if a parser is created as ArgumentParser(default_config_files=['~/.myapp.yaml', '/etc/myapp.yaml']), when parsing if any of those two config files exist it will be parsed and used to override the defaults. All matched config files are parsed and applied in the given order. The default config files are always parsed first, this means that any command line argument will override its values.

It is also possible to add an argument to explicitly provide a configuration file path. Providing a config file as an argument does not disable the parsing of default_config_files. The config argument would be parsed in the specific position among the command line arguments. Therefore the arguments found after would override the values from that config file. The config argument can be given multiple times, each overriding the values of the previous. Using the example parser from the nested-namespaces section above, we could have the following config file in yaml format:

# File: example.yaml
lev1:
  opt1: from yaml 1
  opt2: from yaml 2

Then in python adding a config file argument and parsing some dummy arguments, the following would be observed:

config

cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) with open("example.yaml", "w") as f: f.write("lev1:n opt1: from yaml 1n opt2: from yaml 2n")

config

os.chdir(cwd) shutil.rmtree(tmpdir)

config

>>> from jsonargparse import ArgumentParser >>> parser = ArgumentParser() >>> parser.add_argument("--lev1.opt1", default="from default 1") # doctest: +IGNORE_RESULT >>> parser.add_argument("--lev1.opt2", default="from default 2") # doctest: +IGNORE_RESULT >>> parser.add_argument("--config", action="config") # doctest: +IGNORE_RESULT >>> cfg = parser.parse_args(["--lev1.opt1", "from arg 1", "--config", "example.yaml", "--lev1.opt2", "from arg 2"]) >>> cfg.lev1.opt1 'from yaml 1' >>> cfg.lev1.opt2 'from arg 2'

Instead of providing a path to a configuration file, a string with the configuration content can also be provided.

config

>>> cfg = parser.parse_args(["--config", '{"lev1":{"opt1":"from string 1"}}']) >>> cfg.lev1.opt1 'from string 1'

The config file can also be provided as an environment variable as explained in section environment-variables. The configuration file environment variable is the first one to be parsed. Any other argument provided through an environment variable would override the config file one.

A configuration file or string can also be parsed without parsing command line arguments. The methods for this are :py.ArgumentParser.parse_path and :py.ArgumentParser.parse_string to parse a config file or a config string respectively.

Serialization

Parsers that have an action="config" argument also include a --print_config option. This is useful particularly for command line tools with a large set of options to create an initial config file including all default values. If the ruyaml package is installed, the config can be printed having the help descriptions content as yaml comments by using --print_config=comments. Another option is --print_config=skip_null which skips entries whose value is null.

From within python it is also possible to serialize a config object by using either the :py.ArgumentParser.dump or :py.ArgumentParser.save methods. Three formats with a particular style are supported: yaml, json and json_indented. It is possible to add more dumping formats by using the .set_dumper function. For example to allow dumping using PyYAML's default_flow_style do the following:

import yaml from jsonargparse import set_dumper

def custom_yaml_dump(data):

return yaml.safe_dump(data, default_flow_style=True)

set_dumper("yaml_custom", custom_yaml_dump)

Custom loaders

The yaml parser mode (see :py.ArgumentParser.__init__) uses for loading a subclass of yaml.SafeLoader with two modifications. First, it supports float's scientific notation, e.g. '1e-3' => 0.001 (unlike default PyYAML which considers '1e-3' a string). Second, text within curly braces is considered a string, e.g. '{text}' (unlike default PyYAML which parses this as{'text': None}). It is possible to replace the yaml loader or add a loader as a new parser mode via the :func:`.set_loader` function. For example if you need a custom PyYAML loader it can be registered and used as follows: .. testcode:: import yaml from jsonargparse import ArgumentParser, set_loader class CustomLoader(yaml.SafeLoader): ... def custom_yaml_load(stream): return yaml.load(stream, Loader=CustomLoader) set_loader("yaml_custom", custom_yaml_load) parser = ArgumentParser(parser_mode="yaml_custom") When setting a loader based on a library different from PyYAML, theexceptionsthat it raises when there are failures should be given to :func:`.set_loader`. .. _classes-methods-functions: Classes, methods and functions ============================== It is good practice to write python code in which parameters have type hints and these are described in the docstrings. To make this well written code configurable, it wouldn't make sense to duplicate information of types and parameter descriptions. To avoid this duplication, jsonargparse includes methods to automatically add annotated parameters as arguments, see :class:`.SignatureArguments`. Take for example a class with its init and a method with docstrings as follows: .. testsetup:: class_method sys.argv = ["", "--myclass.init.foo={}", "--myclass.method.bar=0"] class MyBaseClass: pass .. testcode:: class_method from typing import Dict, Union, List class MyClass(MyBaseClass): def __init__(self, foo: Dict[str, Union[int, List[int]]], **kwargs): """Initializer for MyClass. Args: foo: Description for foo. """ super().__init__(**kwargs) ... def mymethod(self, bar: float, baz: bool = False): """Description for mymethod. Args: bar: Description for bar. baz: Description for baz. """ ... BothMyClassandmymethodcan easily be made configurable, the class initialized and the method executed as follows: .. testcode:: class_method from jsonargparse import ArgumentParser parser = ArgumentParser() parser.add_class_arguments(MyClass, "myclass.init") parser.add_method_arguments(MyClass, "mymethod", "myclass.method") cfg = parser.parse_args() myclass = MyClass(**cfg.myclass.init.as_dict()) myclass.mymethod(**cfg.myclass.method.as_dict()) The :func:`add_class_arguments` call adds to themyclass.initkey theitemsargument with description as in the docstring, sets it as required since it lacks a default value. When parsed, it is validated according to the type hint, i.e., a dict with values ints or list of ints. Also since the init has the**kwargsargument, the keyword arguments fromMyBaseClassare also added to the parser. Similarly, the :func:`add_method_arguments` call adds to themyclass.methodkey, the argumentsvalueas a required float andflagas an optional boolean with default value false. Instantiation of several classes added with :func:`add_class_arguments` can be done more simply for an entire config object using :py:meth:`.ArgumentParser.instantiate_classes`. For the example above runningcfg = parser.instantiate_classes(cfg)would result incfg.myclass.initcontaining an instance ofMyClassinitialized with whatever command line arguments were parsed. When parsing from a configuration file (see :ref:`configuration-files`) all the values can be given in a single config file. For convenience it is also possible that the values for each of the argument groups created by the calls to add signatures methods can be parsed from independent files. This means that for the example above there could be one general config file with contents: .. code-block:: yaml myclass: init: myclass.yaml method: mymethod.yaml Then the filesmyclass.yamlandmymethod.yamlwould include the settings for the instantiation of the class and the call to the method respectively. A wide range of type hints are supported for the signature parameters. For exact details go to section :ref:`type-hints`. Some notes about the add signature methods are: - All positional only parameters must have a type, otherwise the add arguments functions raise an exception. - Keyword parameters are ignored if they don't have at least one type that is supported. - Parameters whose name starts with_are considered internal and ignored. - The signature methods have askipparameter which can be used to exclude adding some arguments, e.g.parser.add_method_arguments(MyClass, 'mymethod', skip={'flag'}). .. note:: The signatures support is intended to be non-intrusive. It is by design that there is no need to inherit from a class, add decorators, or use special type hints and default values. This has several advantages. For example it is possible to use classes from third party libraries which is not possible for developers to modify. Docstring parsing ----------------- To get parameter docstrings in the parser help, the `docstring-parser <https://pypi.org/project/docstring-parser/>`__ package is required. This package is included when installing jsonargparse with thesignaturesextras require as explained in section :ref:`installation`. A couple of options can be configured, both related to docstring parsing speed. By default docstrings are parsed used withdocstring_parser.DocstringStyle.AUTO, which means that it is attempted to parse docstrings with all supported styles. If the relevant codebase uses a single style, this is inefficient. A single style can be configured as follows: .. testcode:: docstrings from docstring_parser import DocstringStyle from jsonargparse import set_docstring_parse_options set_docstring_parse_options(style=DocstringStyle.REST) The second option that can be configured is the support for `attribute docstrings <https://peps.python.org/pep-0257/#what-is-a-docstring>`__ (i.e. literal strings in the line after an attribute is defined). By default this feature is disabled and enabling it makes the parsing slower even for classes that don't have attribute docstrings. To enable, do as follows: .. testcode:: docstrings from dataclasses import dataclass from jsonargparse import set_docstring_parse_options set_docstring_parse_options(attribute_docstrings=True) @dataclass class Options: """Options for a competition winner.""" name: str """Name of winner.""" prize: int = 100 """Amount won.""" .. testcleanup:: docstrings set_docstring_parse_options(style=DocstringStyle.GOOGLE) set_docstring_parse_options(attribute_docstrings=False) Classes from functions ---------------------- In some cases there are functions which return an instance of a class. To add this to a parser such that :py:meth:`.ArgumentParser.instantiate_classes` calls this function, the example above would change to: .. testsetup:: class_from_function class MyClass: pass def instantiate_myclass() -> MyClass: return MyClass() .. testcode:: class_from_function from jsonargparse import ArgumentParser, class_from_function parser = ArgumentParser() dynamic_class = class_from_function(instantiate_myclass) parser.add_class_arguments(dynamic_class, "myclass.init") .. note:: :func:`.class_from_function` requires the input function to have a return type annotation that must be the class type it returns. Classes created with :func:`.class_from_function` can be selected usingclass_pathfor :ref:`sub-classes`. For example, if :func:`.class_from_function` is run in a modulemy_moduleas: .. testcode:: class_from_function class_from_function(instantiate_myclass, name="MyClass") Then theclass_pathfor the created class would bemy_module.MyClass. Parameter resolvers ------------------- Three techniques are implemented for resolving signature parameters. One makes use of python's `Abstract Syntax Trees (AST) <https://docs.python.org/3/library/ast.html>`__ library and the second is based on assumptions of class inheritance. The AST resolver is used first and only when AST fails, the assumptions resolver is run as fallback. The third resolver uses stub files*.pyiand is applied on top of both the AST and assumptions resolvers. Unresolved parameters ^^^^^^^^^^^^^^^^^^^^^ The parameter resolvers make a best effort to determine the correct names and types that the parser should accept. However, there can be cases not yet supported or cases for which it would be impossible to support. To somewhat overcome these limitations, there is a special keydict_kwargsthat can be used to provide arguments that will not be validated during parsing, but will be used for class instantiation. It is calleddict_kwargsbecause there are use cases in which**kwargsis used just as a dict, thus it also serves that purpose. Take for example the following parsing and instantiation: .. testsetup:: unresolved sys.argv = ["", "--myclass=MyClass"] class MyClass: def __init__(self, foo: int = 0, **kwargs): super().__init__(**kwargs) ... MyClass.__module__ = "jsonargparse_tests" jsonargparse_tests.MyClass = MyClass .. testcode:: unresolved from jsonargparse import ArgumentParser parser = ArgumentParser() parser.add_argument("--myclass", type=MyClass) cfg = parser.parse_args() cfg_init = parser.instantiate_classes(cfg) IfMyClass.__init__has**kwargswith some unresolved parameters, the following could be a valid config file: .. code-block:: yaml class_path: MyClass init_args: foo: 1 dict_kwargs: bar: 2 The value forbarwill not be validated, but the class will be instantiated asMyClass(foo=1, bar=2). Assumptions resolver ^^^^^^^^^^^^^^^^^^^^ The assumptions resolver only considers classes. Whenever the__init__method hasargs`` and/or ``kwargs``, the resolver assumes that these are directly forwarded to the next parent class, i.e. ``__init__`` includes a line like ``super().__init__(args, **kwargs). Thus, it blindly collects the__init__parameters of parent classes. The collected parameters will be incorrect if the code does not follow this pattern. This is why it is only used as fallback when the AST resolver fails. .. _ast-resolver: AST resolver ^^^^^^^^^^^^ The AST resolver analyzes the source code and tries to figure out how theargs`` and ``kwargs`` are used to further find more accepted parameters. This type of resolving is limited to a few specific cases since there are endless possibilities for what code can do. The supported cases are illustrated below. Bear in mind that the code does not need to be exactly like this. The important detail is how ``argsand**kwargsare used, not other parameters, or the names of variables, or the complexity of the code that is unrelated to these variables. .. testsetup:: ast_resolver class BaseClass: pass class SomeClass: def __init__(self, **kwargs): pass class ChildClass(BaseClass): def __init__(self, *args, **kwargs): pass **Cases for statements in functions or methods** .. testcode:: ast_resolver def calls_a_function(*args, **kwargs): a_function(*args, **kwargs) def calls_a_method(*args, **kwargs): an_instance = SomeClass() an_instance.a_method(*args, **kwargs) def calls_a_static_method(*args, **kwargs): an_instance = SomeClass() an_instance.a_static_method(*args, **kwargs) def calls_a_class_method(*args, **kwargs): SomeClass.a_class_method(*args, **kwargs) def calls_local_import(**kwargs): import some_module some_module.a_callable(**kwargs) def pops_from_kwargs(**kwargs): val = kwargs.pop("name", "default") def gets_from_kwargs(**kwargs): val = kwargs.get("name", "default") def constant_conditional(**kwargs): if global_boolean_1: first_function(**kwargs) elif not global_boolean_2: second_function(**kwargs) else: third_function(**kwargs) **Cases for classes** .. testcode:: ast_resolver class PassThrough(BaseClass): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) class CallMethod: def __init__(self, *args, **kwargs): self.a_method(*args, **kwargs) class AttributeUseInMethod: def __init__(self, **kwargs): self._kwargs = kwargs def a_method(self): a_callable(**self._kwargs) class AttributeUseInProperty: def __init__(self, **kwargs): self._kwargs = kwargs @property def a_property(self): return a_callable(**self._kwargs) class DictUpdateUseInMethod: def __init__(self, **kwargs): self._kwargs = dict(p1=1) # Can also be: self._kwargs = {'p1': 1} self._kwargs.update(**kwargs) # Can also be: self._kwargs = dict(p1=1, **kwargs) def a_method(self): a_callable(**self._kwargs) class InstanceInClassmethod: @classmethod def get_instance(cls, **kwargs): return cls(**kwargs) class NonImmediateSuper(BaseClass): def __init__(self, *args, **kwargs): super(BaseClass, self).__init__(*args, **kwargs) **Cases for class instance defaults** .. testcode:: ast_resolver # Class instance: only keyword arguments withast.Constant` value class_instance: SomeClass = SomeClass(param=1)

# Lambda returning class instance: only keyword arguments with `ast.Constant value class_instance: Callable[[type], BaseClass] = lambda a: ChildClass(a, param=2.3)

There can be other parameters apart from *args and **kwargs, thus in the cases above, the signatures can be for example like name(p1: int, k1: str = 'a', **kws). Also when internally calling some function or instantiating a class, there can be additional parameters. For example in:

def calls_a_function(args,kwargs): a_function(args, param=1, **kwargs)

The param parameter would be excluded from the resolved parameters because it is internally hard coded.

A special case which is supported but with caveats, is multiple calls that use **kwargs. For example:

ast_resolver

def conditional_calls(kwargs): if condition_1: first_function(kwargs)
elif condition_2:

second_function(**kwargs)

else:

third_function(**kwargs)

The resolved parameters that have the same type hint and default across all calls are supported normally. When there is a discrepancy between the calls, the parameters behave differently and are shown in the help with the default like Conditional<ast-resolver> {DEFAULT_1, ...}. The main difference is that these parameters are not included in :py.ArgumentParser.get_defaults or the output of --print_config. This is necessary because the parser does not know which of the calls will be used at runtime, and adding them would cause :py.ArgumentParser.instantiate_classes to fail due to unexpected keyword arguments.

Note

The parameter resolvers log messages of failures and unsupported cases. To view these logs, set the environment variable JSONARGPARSE_DEBUG to a non-empty truthy value. The supported cases are limited and it is highly encouraged that people create issues requesting the support for new ones. However, note that when a case is highly convoluted it could be a symptom that the respective code is in need of refactoring.

Stubs resolver

The stubs resolver makes use of the typeshed-client package to identify parameters and their type hints from stub files *.pyi. To enable this resolver, install jsonargparse with the signatures extras require as explained in section installation.

Many of the types defined in stub files use the latest syntax for type hints, that is, bitwise or operator | for unions and generics, e.g. list[<type>] instead of typing.List[<type>], see PEPs 604 and 585. On python>=3.10 these are fully supported. On python<=3.9 backporting these types is attempted and in some cases it can fail. On failure the type annotation is set to Any.

Most of the types in the Python standard library have their types in stubs. An example from the standard library would be:

stubs_resolver

>>> from random import uniform

>>> parser = ArgumentParser() >>> parser.add_function_arguments(uniform, "uniform") # doctest: +IGNORE_RESULT >>> parser.parse_args(["--uniform.a=0.7", "--uniform.b=3.4"]) Namespace(uniform=Namespace(a=0.7, b=3.4))

Without the stubs resolver, the :py.SignatureArguments.add_function_arguments call requires the fail_untyped=False option. This has the disadvantage that type Any is given to the a and b arguments, instead of float. And this means that the parser would not fail if given an invalid value, for instance a string.

It is not possible to know the defaults of parameters discovered only because of the stubs. In these cases in the parser help the default is shown as Unknown<stubs-resolver> and not included in :py.ArgumentParser.get_defaults or the output of --print_config.

Class type and sub-classes

It is possible to use an arbitrary class as a type such that the argument accepts an instance of this class or any derived subclass. This practice is known as dependency injection. In the config file a class is represented by a dictionary with a class_path entry indicating the dot notation expression to import the class, and optionally some init_args that would be used to instantiate it. When parsing, it will be checked that the class can be imported, that it is a subclass of the given type and that init_args values correspond to valid arguments to instantiate it. After parsing, the config object will include the class_path and init_args entries. To get a config object with all sub-classes instantiated, the :py.ArgumentParser.instantiate_classes method is used. The skip parameter of the signature methods can also be used to exclude arguments within subclasses. This is done by giving its relative destination key, i.e. as param.init_args.subparam.

A simple example would be having some config file config.yaml as:

myclass:
  calendar:
    class_path: calendar.Calendar
    init_args:
      firstweekday: 1

Then in python:

subclasses

cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) with open("config.yaml", "w") as f: f.write("myclass:n calendar:n class_path: calendar.Calendarn init_args:n firstweekday: 1n")

subclasses

os.chdir(cwd) shutil.rmtree(tmpdir)

subclasses

>>> from calendar import Calendar

>>> class MyClass: ... def __init__(self, calendar: Calendar): ... self.calendar = calendar ...

>>> parser = ArgumentParser() >>> parser.add_class_arguments(MyClass, "myclass") # doctest: +IGNORE_RESULT

>>> cfg = parser.parse_path("config.yaml") >>> cfg.myclass.calendar.as_dict() {'class_path': 'calendar.Calendar', 'init_args': {'firstweekday': 1}}

>>> cfg = parser.instantiate_classes(cfg) >>> cfg.myclass.calendar.getfirstweekday() 1

In this example the class_path points to the same class used for the type. But a subclass of Calendar with an extended set of init parameters would also work.

An individual argument can also be added having as type a class, i.e. parser.add_argument('--calendar', type=Calendar). There is also another method :py.SignatureArguments.add_subclass_arguments which does the same as add_argument, but has some added benefits: 1) the argument is added in a new group automatically; 2) the argument values can be given in an independent config file by specifying a path to it; and 3) by default sets a useful metavar and help strings.

Note

Classes will be parsed and instantiated when given as value a dict with class_path and init_args if the corresponding parameter has type Any, or when fail_untyped=False which defaults to type Any.

Note

It is also possible to provide to class_path a function that has as return type a class. The accepted init_args would be the parameters of that function.

Command line

The help of the parser does not show details for a type class since this depends on the subclass. To get details for a particular subclass there is a specific help option that receives the import path. Take for example a parser defined as:

from calendar import Calendar from jsonargparse import ArgumentParser

parser = ArgumentParser() parser.add_argument("--calendar", type=Calendar)

The help for a corresponding subclass could be printed as:

python tool.py --calendar.help calendar.TextCalendar

In the command line, a subclass can be specified through multiple command line arguments:

python tool.py \
  --calendar.class_path calendar.TextCalendar \
  --calendar.init_args.firstweekday 1

For convenience, the arguments can be somewhat shorter by omitting .class_path and .init_args and only specifying the name of the subclass instead of the full import path.

python tool.py --calendar TextCalendar --calendar.firstweekday 1

Specifying the name of the subclass works for subclasses in modules that have been imported before parsing. Abstract classes and private classes (module or name starting with '_') are not considered. All the subclasses resolvable by its name can be seen in the general help python tool.py --help.

When the base class is not abstract, the class_path can be omitted, by giving directly init_args, for example:

python tool.py --calendar.firstweekday 2

would implicitly use calendar.Calendar as the class path.

Default values

For a parameter that has a class as type, it might also be wanted to set a default value for it. Special care must be taken when doing this, could be considered bad practice and be a good idea to avoid in most cases. The issue is that classes are normally mutable. Depending on how the parameter value is used, its default class instance in the signature could be changed. This goes against what a default value is expected to be and lead to bugs which are difficult to debug.

Since there are some legitimate use cases for class instances in defaults, they are supported with a particular behavior and recommendations. An example is:

instance_default

class MyClass:
def __init__(

self, calendar: Calendar = Calendar(firstweekday=1),

):

self.calendar = calendar

Adding this class to a parser will work without issues. The ast-resolver in limited cases determines how to instantiate the original default. The parsing methods would provide a dict with class_path and init_args instead of the class instance. Furthermore, if :py.ArgumentParser.instantiate_classes is used, a new instance of the class is created, thereby avoiding issues related to the mutability of the default.

Since the ast-resolver only supports limited cases, or when the source code is not available, a second approach is to use the special function .lazy_instance to instantiate the default. Continuing with the same example above, this would be:

instance_default

from jsonargparse import lazy_instance

class MyClass:
def __init__(

self, calendar: Calendar = lazy_instance(Calendar, firstweekday=1),

):

self.calendar = calendar

Like this, the parsed default will be a dict with class_path and init_args, again avoiding the risk of mutability.

Note

In python there can be some classes or functions for which it is not possible to determine its import path from the object alone. When using one of these as a default would cause a failure when serializing because what gets saved in the config file is the import path. To overcome this problem use the .register_unresolvable_import_paths function giving it the module from where the respective object can be imported.

Argument linking

Some use cases could require adding arguments from multiple classes and some parameters get a value automatically computed from other arguments. This behavior can be obtained by using the :py.ArgumentLinking.link_arguments method.

There are two types of links, defined with apply_on='parse' or apply_on='instantiate'. As the names suggest, the former are set when calling one of the parse methods and the latter are set when calling :py.ArgumentParser.instantiate_classes.

Applied on parse

For parsing links, source keys can be individual arguments or nested groups. The target key has to be a single argument. The keys can be inside init_args of a subclass. The compute function should accept as many positional arguments as there are sources and return a value of type compatible with the target. An example would be the following:

class Model:
def __init__(self, batch_size: int):

self.batch_size = batch_size

class Data:
def __init__(self, batch_size: int = 5):

self.batch_size = batch_size

parser = ArgumentParser() parser.add_class_arguments(Model, "model") parser.add_class_arguments(Data, "data") parser.link_arguments("data.batch_size", "model.batch_size", apply_on="parse")

As argument and in config files only data.batch_size should be specified. Then whatever value it has will be propagated to model.batch_size.

An example of a target being in a subclass is:

class Logger:
def __init__(self, save_dir: Optional[str] = None):

self.save_dir = save_dir

class Trainer:
def __init__(

self, save_dir: Optional[str] = None, logger: Union[bool, Logger, List[Logger]] = False,

):

self.logger = logger

parser = ArgumentParser() parser.add_class_arguments(Trainer, "trainer") parser.link_arguments("trainer.save_dir", "trainer.logger.init_args.save_dir")

The link gets applied to the logger parameter when it is a single subclass and applied to all elements of a list of subclasses. If a subclass does not define the targeted init_args parameter, the link is ignored.

Applied on instantiate

For instantiation links, sources can be class groups (added with :py.SignatureArguments.add_class_arguments) or subclass arguments (see sub-classes). The source key can be the entire instantiated object or an attribute of the object. The target key has to be a single argument and can be inside init_args of a subclass. The order of instantiation used by :py.ArgumentParser.instantiate_classes is automatically determined based on the links. The set of all instantiation links must be a directed acyclic graph. An example would be the following:

class Model:
def __init__(self, num_classes: int):

self.num_classes = num_classes

class Data:
def __init__(self):

self.num_classes = get_num_classes()

parser = ArgumentParser() parser.add_class_arguments(Model, "model") parser.add_class_arguments(Data, "data") parser.link_arguments("data.num_classes", "model.num_classes", apply_on="instantiate")

This link would imply that :py.ArgumentParser.instantiate_classes instantiates Data first, then use the num_classes attribute to instantiate Model.

Variable interpolation

One of the possible reasons to add a parser mode (see custom-loaders) can be to have support for variable interpolation in yaml files. Any library could be used to implement a loader and configure a mode for it. Without needing to implement a loader function, an omegaconf parser mode is available out of the box when this package is installed.

Take for example a yaml file as:

server:
  host: localhost
  port: 80
client:
  url: http://${server.host}:${server.port}/

omegaconf

example = """ server: host: localhost port: 80 client: url: http://${server.host}:${server.port}/ """ cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) with open("example.yaml", "w") as f: f.write(example)

omegaconf

os.chdir(cwd) shutil.rmtree(tmpdir)

This yaml could be parsed as follows:

omegaconf

>>> @dataclass ... class ServerOptions: ... host: str ... port: int ...

>>> @dataclass ... class ClientOptions: ... url: str ...

>>> parser = ArgumentParser(parser_mode="omegaconf") >>> parser.add_argument("--server", type=ServerOptions) # doctest: +IGNORE_RESULT >>> parser.add_argument("--client", type=ClientOptions) # doctest: +IGNORE_RESULT >>> parser.add_argument("--config", action="config") # doctest: +IGNORE_RESULT

>>> cfg = parser.parse_args(["--config=example.yaml"]) >>> cfg.client.url 'http://localhost:80/'

Note

The parser_mode='omegaconf' provides support for OmegaConf's variable interpolation in a single yaml file. It is not possible to do interpolation across multiple yaml files or in an isolated individual command line argument.

Environment variables

The jsonargparse parsers can also get values from environment variables. The parser checks existing environment variables whose name is of the form [PREFIX_][LEV__]*OPT, that is, all in upper case, first a prefix (set by env_prefix, or if unset the prog without extension or none if set to False) followed by underscore and then the argument name replacing dots with two underscores. Using the parser from the nested-namespaces section above, in your shell you would set the environment variables as:

export APP_LEV1__OPT1='from env 1'
export APP_LEV1__OPT2='from env 2'

Then in python the parser would use these variables, unless overridden by the command line arguments, that is:

env

os.environ["APP_LEV1__OPT1"] = "from env 1" os.environ["APP_LEV1__OPT2"] = "from env 2"

env

>>> parser = ArgumentParser(env_prefix="APP", default_env=True) >>> parser.add_argument("--lev1.opt1", default="from default 1") # doctest: +IGNORE_RESULT >>> parser.add_argument("--lev1.opt2", default="from default 2") # doctest: +IGNORE_RESULT >>> cfg = parser.parse_args(["--lev1.opt1", "from arg 1"]) >>> cfg.lev1.opt1 'from arg 1' >>> cfg.lev1.opt2 'from env 2'

Note that when creating the parser, default_env=True was given. By default :py.ArgumentParser.parse_args does not parse environment variables. If default_env is left unset, environment variable parsing can also be enabled by setting in your shell JSONARGPARSE_DEFAULT_ENV=true.

There is also the :py.ArgumentParser.parse_env function to only parse environment variables, which might be useful for some use cases in which there is no command line call involved.

If a parser includes an action="config" argument, then the environment variable for this config file will be parsed before all the other environment variables.

Sub-commands

A way to define parsers in a modular way is what in argparse is known as sub-commands. However, to promote modularity, in jsonargparse sub-commands work a bit different than in argparse. To add sub-commands to a parser, the :py.ArgumentParser.add_subcommands method is used. Then an existing parser is added as a sub-command using .add_subcommand. In a parsed config object the sub-command will be stored in the subcommand entry (or whatever dest was set to), and the values of the sub-command will be in an entry with the same name as the respective sub-command. An example of defining a parser with sub-commands is the following:

from jsonargparse import ArgumentParser

... parser_subcomm1 = ArgumentParser() parser_subcomm1.add_argument("--op1") ... parser_subcomm2 = ArgumentParser() parser_subcomm2.add_argument("--op2") ... parser = ArgumentParser(prog="app") parser.add_argument("--op0") subcommands = parser.add_subcommands() subcommands.add_subcommand("subcomm1", parser_subcomm1) subcommands.add_subcommand("subcomm2", parser_subcomm2)

Then some examples of parsing are the following:

>>> parser.parse_args(["subcomm1", "--op1", "val1"]) # doctest: +IGNORE_RESULT Namespace(op0=None, subcommand='subcomm1', subcomm1=Namespace(op1='val1')) >>> parser.parse_args(["--op0", "val0", "subcomm2", "--op2", "val2"]) # doctest: +IGNORE_RESULT Namespace(op0='val0', subcommand='subcomm2', subcomm2=Namespace(op2='val2'))

Parsing config files with :py.ArgumentParser.parse_path or :py.ArgumentParser.parse_string is also possible. The config file is not required to specify a value for subcommand. For the example parser above a valid yaml would be:

# File: example.yaml
op0: val0
subcomm1:
  op1: val1

Parsing of environment variables works similar to .ActionParser. For the example parser above, all environment variables for subcomm1 would have as prefix APP_SUBCOMM1_ and likewise for subcomm2 as prefix APP_SUBCOMM2_. The sub-command to use could be chosen by setting environment variable APP_SUBCOMMAND.

It is possible to have multiple levels of sub-commands. With multiple levels there is one basic requirement: the sub-commands must be added in the order of the levels. This is, first call add_subcommands and add_subcommand for the first level. Only after do the same for the second level, and so on.

Json schemas

The .ActionJsonSchema class is provided to allow parsing and validation of values using a json schema. This class requires the jsonschema python package. Though note that jsonschema is not a requirement of the minimal jsonargparse install. To enable this functionality install with the jsonschema extras require as explained in section installation.

Check out the jsonschema documentation to learn how to write a schema. The current version of jsonargparse uses Draft7Validator. Parsing an argument using a json schema is done like in the following example:

>>> from jsonargparse import ActionJsonSchema

>>> schema = { ... "type": "object", ... "properties": { ... "price": {"type": "number"}, ... "name": {"type": "string"}, ... }, ... }

>>> parser = ArgumentParser() >>> parser.add_argument("--json", action=ActionJsonSchema(schema=schema)) # doctest: +IGNORE_RESULT

>>> parser.parse_args(["--json", '{"price": 1.5, "name": "cookie"}']) Namespace(json={'price': 1.5, 'name': 'cookie'})

Instead of giving a json string as argument value, it is also possible to provide a path to a json/yaml file, which would be loaded and validated against the schema. If the schema defines default values, these will be used by the parser to initialize the config values that are not specified. When adding an argument with the .ActionJsonSchema action, you can use "%s" in the help string so that in that position the schema is printed.

Jsonnet files

The Jsonnet support requires jsonschema and jsonnet python packages which are not included with minimal jsonargparse install. To enable this functionality install jsonargparse with the jsonnet extras require as explained in section installation.

By default an .ArgumentParser parses configuration files as yaml. However, if instantiated giving parser_mode='jsonnet', then parse_args, parse_path and parse_string will expect config files to be in jsonnet format instead. Example:

jsonnet

cwd = os.getcwd() tmpdir = tempfile.mkdtemp(prefix="_jsonargparse_doctest") os.chdir(tmpdir) with open("example.jsonnet", "w") as f: f.write("{}n")

jsonnet

os.chdir(cwd) shutil.rmtree(tmpdir)

jsonnet

from jsonargparse import ArgumentParser

parser = ArgumentParser(parser_mode="jsonnet") parser.add_argument("--config", action="config") cfg = parser.parse_args(["--config", "example.jsonnet"])

Jsonnet files are commonly parametrized, thus requiring external variables for parsing. For these cases, instead of changing the parser mode away from yaml, the .ActionJsonnet class can be used. This action allows to define an argument which would be a jsonnet string or a path to a jsonnet file. Moreover, another argument can be specified as the source for any external variables required, which would be either a path to or a string containing a json dictionary of variables. Its use would be as follows:

jsonnet

from jsonargparse import ArgumentParser, ActionJsonnet

parser = ArgumentParser() parser.add_argument("--in_ext_vars", type=dict) parser.add_argument("--in_jsonnet", action=ActionJsonnet(ext_vars="in_ext_vars"))

For example, if a jsonnet file required some external variable param, then the jsonnet and the external variable could be given as:

jsonnet

cfg = parser.parse_args(["--in_ext_vars", '{"param": 123}', "--in_jsonnet", "example.jsonnet"])

Note that the external variables argument must be provided before the jsonnet path so that this dictionary already exists when parsing the jsonnet.

The .ActionJsonnet class also accepts as argument a json schema, in which case the jsonnet would be validated against this schema right after parsing.

Parsers as arguments

Sometimes it is useful to take an already existing parser that is required standalone in some part of the code, and reuse it to parse an inner node of another more complex parser. For these cases an argument can be defined using the .ActionParser class. An example of how to use this class is the following:

from jsonargparse import ArgumentParser, ActionParser

inner_parser = ArgumentParser(prog="app1") inner_parser.add_argument("--op1") ... outer_parser = ArgumentParser(prog="app2") outer_parser.add_argument("--inner.node", title="Inner node title", action=ActionParser(parser=inner_parser))

When using the .ActionParser class, the value of the node in a config file can be either the complex node itself, or the path to a file which will be loaded and parsed with the corresponding inner parser. Naturally using action="config" to parse a complete config file will parse the inner nodes correctly.

Note that when adding inner_parser a title was given. In the help, the added parsers are shown as independent groups starting with the given title. It is also possible to provide a description.

Regarding environment variables, the prefix of the outer parser will be used to populate the leaf nodes of the inner parser. In the example above, if inner_parser is used to parse environment variables, then as normal APP1_OP1 would be checked to populate option op1. But if outer_parser is used, then APP2_INNER__NODE__OP1 would be checked to populate inner.node.op1.

An important detail to note is that the parsers that are given to .ActionParser are internally modified. Therefore, to use the parser both as standalone and as inner node, it is necessary to implement a function that instantiates the parser. This function would be used in one place to get an instance of the parser for standalone parsing, and in some other place use the function to provide an instance of the parser to .ActionParser.

Tab completion

Tab completion is available for jsonargparse parsers by using the argcomplete package. There is no need to implement completer functions or to call argcomplete.autocomplete since this is done automatically by :py.ArgumentParser.parse_args. The only requirement to enable tab completion is to install argcomplete either directly or by installing jsonargparse with the argcomplete extras require as explained in section installation. Then the tab completion can be enabled globally for all argcomplete compatible tools or for each individual tool. A simple example.py tool would be:

tab_completion

sys.argv = [""]

tab_completion

#!/usr/bin/env python3

from typing import Optional from jsonargparse import ArgumentParser

parser = ArgumentParser() parser.add_argument("--bool", type=Optional[bool])

parser.parse_args()

Then in a bash shell you can add the executable bit to the script, activate tab completion and use it as follows:

$ chmod +x example.py
$ eval "$(register-python-argcomplete example.py)"

$ ./example.py --bool <TAB><TAB>
false  null   true
$ ./example.py --bool f<TAB>
$ ./example.py --bool false

Troubleshooting and logging

The standard behavior for the parse methods, when they fail, is to print a short message and terminate the process with a non-zero exit code. This is problematic during development since there is not enough information to track down the root of the problem. Without the need to change the source code, this default behavior can be changed such that in case of failure, a ParseError exception is raised and the full stack trace is printed. This is done by setting the JSONARGPARSE_DEBUG environment variable to a non-empty truthy value.

The parsers from jsonargparse log some basic events, though by default this is disabled. To enable, the logger argument should be set when creating an .ArgumentParser object. The intended use is to give as value an already existing logger object which is used for the whole application. For convenience, to enable a default logger the logger argument can also receive True or a string which sets the name of the logger or a dictionary that can include the name and the level, e.g. {"name": "myapp", "level": "ERROR"}. If reconplogger is installed, setting logger to True or a dictionary without specifying a name, then the reconplogger is used. If reconplogger is installed and the JSONARGPARSE_DEBUG environment variable is set, then the logging level becomes DEBUG.