Skip to main content

Parsing of command line options, yaml/jsonnet config files and/or environment variables based on argparse.

Project description


This package is an extension to python’s argparse which simplifies parsing of configuration options from command line arguments, json configuration files (yaml or jsonnet supersets), environment variables and hard-coded defaults.

The aim is similar to other projects such as configargparse, yconf, confuse, typer, OmegaConf, Fire and click. The obvious question is, why yet another package similar to many already existing ones? The answer is simply that none of the existing projects had the exact features we wanted and after analyzing the alternatives it seemed simpler to start a new project.


  • Great support of type hint annotations for automatic creation of parsers and minimal boilerplate command line interfaces.
  • Support for nested namespaces which makes it possible to parse config files with non-flat hierarchies.
  • Parsing of relative paths within config files and path lists.
  • Support for two popular supersets of json: yaml and jsonnet.
  • Parsers can be configured just like with python’s argparse, thus it has a gentle learning curve.
  • Several convenient types and action classes to ease common parsing use cases (paths, comparison operators, json schemas, enums, regex strings, …).
  • Support for command line tab argument completion using argcomplete.
  • Configuration values are overridden based on the following precedence.
    • Parsing command line: command line arguments (might include config files) > environment variables > default config files > defaults.
    • Parsing files: config file > environment variables > default config files > defaults.
    • Parsing environment: environment variables > default config files > defaults.


You can install using pip as:

pip install jsonargparse

By default the only dependency that jsonargparse installs is PyYAML. However, several optional features can be enabled by specifying any of the following extras requires: signatures, jsonschema, jsonnet, urls, argcomplete and reconplogger. There is also the all extras require to enable all optional features. Installing jsonargparse with extras require is as follows:

pip install "jsonargparse[signatures,urls]"  # Enable signatures and URLs features
pip install "jsonargparse[all]"              # Enable all optional features

The following table references sections that describe optional features and the corresponding extras requires that enables them.

  urls/fsspec argcomplete jsonnet jsonschema signatures

Basic usage

There are multiple ways of using jsonargparse. The most simple way which requires to write the least amount of code is by using the .CLI function, for example:

from jsonargparse import CLI

def command(
    name: str,
    prize: int = 100
        name: Name of winner.
        prize: Amount won.
    print(f'{name} won {prize}€!')

if __name__ == '__main__':

Then in a shell you could run:

$ python Lucky --prize=1000
Lucky won 1000€!

.CLI without arguments searches for functions and classes defined in the same module and in the local context where .CLI is called. Giving a single or a list functions/classes as first argument to .CLI skips the automatic search and only includes what is given.

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

from random import randint
from jsonargparse import CLI

class Main:
    def __init__(
        max_prize: int = 100
            max_prize: Maximum prize that can be awarded.
        self.max_prize = max_prize

    def person(
        name: str
            name: Name of winner.
        return f'{name} won {randint(0, self.max_prize)}€!'

if __name__ == '__main__':

Then in a shell you could run:

$ python --max_prize=1000 person Lucky
Lucky won 632€!

If more than one function is given to .CLI, then any of them can be executed via sub-commands similar to the single class example above, i.e. 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. class [init_arguments] method [arguments]. For more details about the automatic adding of arguments from classes and functions and the use of configuration files refer to section classes-methods-functions.

This simple way of usage 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 required to have type hints, and the values will be validated according to these. Third, the return values of the functions are not automatically printed. .CLI returns its value and it is up to the developer to decide what to do with it. Finally, jsonargparse has many features designed to help in creating convenient argument parsers such as: nested-namespaces, configuration-files, additional type hints (parsing-paths, restricted-numbers, restricted-strings) and much more.

Section parsers explains how to create an argument parser in a low level argparse-style. However, as parsers get more complex, being able to define them in a modular way becomes important. Three mechanisms are available for modularity, see respective sections classes-methods-functions, sub-commands and parser-arguments.



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(
    description='Description for my app.'

    help='Help for option 1.'

    help='Help for option 2.'

After creating the parser, you can use it to parse command line arguments with the .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 error_handler=None in which case a .ParserError is raised.

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__':

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)


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

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')
>>> parser.add_argument('--lev1.opt2', default='from default 2')
>>> 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

class Level1Options:
    """Level 1 options
        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 .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 :.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
  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:

>>> from jsonargparse import ArgumentParser, ActionConfigFile
>>> parser = ArgumentParser()
>>> parser.add_argument('--lev1.opt1', default='from default 1')
>>> parser.add_argument('--lev1.opt2', default='from default 2')
>>> parser.add_argument('--config', action=ActionConfigFile)
>>> 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.

>>> 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 .ArgumentParser.parse_path and .ArgumentParser.parse_string to parse a config file or a config string respectively.


Parsers that have an .ActionConfigFile 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 .ArgumentParser.dump or 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 .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 .set_loader function. For example if you need a custom PyYAML loader it can be registered and used as follows:

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, the exceptions that it raises when there are failures should be given to .set_loader.

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) 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:

>>> parser = ArgumentParser(env_prefix='APP', default_env=True)
>>> parser.add_argument('--lev1.opt1', default='from default 1')
>>> parser.add_argument('--lev1.opt2', default='from default 2')
>>> 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 .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 .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 .ActionConfigFile argument, then the environment variable for this config file will be parsed before all the other environment variables.

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 .SignatureArguments.

Take for example a class with its init and a method with docstrings as follows:

from typing import Dict, Union, List

class MyClass(MyBaseClass):
    def __init__(self, foo: Dict[str, Union[int, List[int]]], **kwargs):
        """Initializer for MyClass.

            foo: Description for foo.

    def mymethod(self, bar: float, baz: bool = False):
        """Description for mymethod.

            bar: Description for bar.
            baz: Description for baz.

Both MyClass and mymethod can easily be made configurable, the class initialized and the method executed as follows:

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())

The add_class_arguments call adds to the myclass.init key the items argument 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 **kwargs argument, the keyword arguments from MyBaseClass are also added to the parser. Similarly, the add_method_arguments call adds to the myclass.method key, the arguments value as a required float and flag as an optional boolean with default value false.

Instantiation of several classes added with add_class_arguments can be done more simply for an entire config object using .ArgumentParser.instantiate_classes. For the example above running cfg = parser.instantiate_classes(cfg) would result in cfg.myclass.init containing an instance of MyClass initialized with whatever command line arguments were parsed.

When parsing from a configuration file (see 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:

  init: myclass.yaml
  method: mymethod.yaml

Then the files myclass.yaml and mymethod.yaml would 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 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 a skip parameter which can be used to exclude adding some arguments, e.g. parser.add_method_arguments(MyClass, 'mymethod', skip={'flag'}).


To get parameter docstrings in the parser help, the docstring-parser package is required. This package is included when installing jsonargparse with the signatures extras require as explained in section installation.

Classes from functions

In some cases there are functions which return an instance of a class. To add this to a parser such that .ArgumentParser.instantiate_classes calls this function, the example above would change to:

from jsonargparse import ArgumentParser, class_from_function

parser = ArgumentParser()
dynamic_class = class_from_function(instantiate_myclass)
parser.add_class_arguments(dynamic_class, 'myclass.init')


.class_from_function requires the input function to have a return type annotation that must be the class type it returns.

Parameter resolvers

Two techniques are implemented for resolving signature parameters. One makes use of python’s Abstract Syntax Trees (AST) library and the other is based on assumptions of class inheritance. The AST resolver is used first and only when AST fails, the assumptions resolver is used as fallback.


To debug issues related to parameter resolving it is recommended to install the reconplogger package and set the JSONARGPARSE_DEBUG before running the code, see logging.

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 key dict_kwargs that can be used to provide arguments that will not be validated during parsing, but will be used for class instantiation. It is called dict_kwargs because there are use cases in which **kwargs is used just as a dict, thus it also serves that purpose.

Take for example the following parsing and instantiation:

from jsonargparse import ArgumentParser

parser = ArgumentParser()
parser.add_argument('--myclass', type=MyClass)
cfg = parser.parse_args()
cfg_init = parser.instantiate_classes(cfg)

If MyClass.__init__ has **kwargs with some unresolved parameters, the following could be a valid config file:

class_path: MyClass
  foo: 1
  bar: 2

The value for bar will not be validated, but the class will be instantiated as MyClass(foo=1, bar=2).

Assumptions resolver

The assumptions resolver only considers classes. Whenever the __init__ method has *args 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

The AST resolver analyzes the source code and tries to figure out how the *args 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 *args and **kwargs are used, not other parameters, or the names of variables, or the complexity of the code that is unrelated to these variables.

# Cases for functions

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)

# Cases for classes

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 CallCallable:
    def __init__(self, *args, **kwargs):
        a_callable(*args, **kwargs)

class AttributeUseInMethod:
    def __init__(self, **kwargs):
        self._kwargs = kwargs

    def a_method(self):

class DictUpdateUseInMethod:
    def __init__(self, **kwargs):
        self._kwargs = dict(p1=1)
        # Could also be: self._kwargs = dict(p1=1, **kwargs)

    def a_method(self):

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 also 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.


The supported cases are limited and it is highly encouraged that people create issues requesting the support for new cases. However, note that when a case is highly convoluted it could be a symptom that the respective code is in need of refactoring.

Argument linking

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

There are two types of links each defined with apply_on='parse' and 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 .ArgumentParser.instantiate_classes.

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.

For instantiation links, only a single source key is supported. The key can be for a class group created using .SignatureArguments.add_class_arguments or a subclass action created using .SignatureArguments.add_subclass_arguments. If the key is only the class group or subclass action, then a compute function is required which takes the source class instance and returns the value to set in target. Alternatively the key can specify a class attribute. The target key has to be a single argument and can be inside init_args of a subclass. The order of instantiation used by .ArgumentParser.instantiate_classes is automatically determined based on the links. The 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 .ArgumentParser.instantiate_classes instantiates Data first, then use the num_classes attribute to instantiate Model.

Type hints

As explained in section classes-methods-functions type hints are required to automatically add arguments from signatures to a parser. Additional to this feature, a type hint can also be used independently when adding a single argument to the parser. 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 support of type hints is designed to not require developers to change their types or default values. In other words, the idea is to support type hints whatever they may be, as opposed to requiring jsonargparse specific types. The types included in jsonargparse.typing are completely generic and could even be useful independent of the argument parsers.

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.
  • Fully supported types are: str, bool, int, float, complex, List, Iterable, Sequence, Any, Union, Optional, Type, Enum, 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, and MutableMapping are supported but only with str or int keys.
  • 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.
  • dataclasses are supported as a type but only for pure data classes and not nested in a type. By pure it is meant that the class only inherits from data classes. Not a mixture of normal classes and data classes. Data classes as fields of other data classes is supported.
  • 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.
  • 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 specifying a class that once instantiated is callable. Running .ArgumentParser.instantiate_classes will instantiate the callable classes. Currently the callable’s arguments and return types are ignored and not validated.

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:

>>> parser.add_argument('--list', type=List[int])
>>> 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
- 1
# config2.yaml
- 2
- 3

Appending works for any type for the list elements. List elements with class type is also supported and the short notation for init_args when used (see sub-classes), gets applied to the last element of the list.

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)
# either int larger than zero or 'off' string
def int_or_off(x): return x if x == 'off' else PositiveInt(x)
parser.add_argument('--op3', type=int_or_off)

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)

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:

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

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)


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.

Class type and sub-classes

It is also possible to use an arbitrary class as a type such that the argument accepts this class or any derived subclass. 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 .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:

    class_path: calendar.Calendar
      firstweekday: 1

Then in python:

>>> from calendar import Calendar

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

>>> parser = ArgumentParser()
>>> parser.add_class_arguments(MyClass, 'myclass')

>>> 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()

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 .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.

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 calendar.TextCalendar

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

python \
  --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 --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 --help.

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. The first approach is using a normal class instance, for example:

class MyClass:
    def __init__(
        calendar: Calendar = Calendar(firstweekday=1),
        self.calendar = calendar

Adding this class to a parser will work without issues. Parsing would also work and if not overridden the default class instance will be found in the respective key of the config object. If --print_config is used, the class instance is just cast to a string. This means that the generated config file must be modified to become a valid input to the parser. Due to the limitations and the risk of mutable default this approach is discouraged.

The second approach which is the recommended one is to use the special function .lazy_instance to instantiate the default. Continuing with the same example above this would be:

from jsonargparse import lazy_instance

class MyClass:
    def __init__(
        calendar: Calendar = lazy_instance(Calendar, firstweekday=1),
        self.calendar = calendar

In this case the default value will still be an instance of Calendar. The difference is that the parsing methods would provide a dict with class_path and init_args instead of the class instance. Furthermore, if .ArgumentParser.instantiate_classes is used a new instance of the class is created thereby avoiding issues related to the mutability of the default.


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.

Final classes

When a class is decorated with .final there shouldn’t be any derived subclass. Using a final class as a type hint works similar to subclasses. The difference is that the init args are given directly in a dictionary without specifying a class_path. Therefore, the code below would accept the corresponding yaml structure.

from jsonargparse.typing import final

class FinalCalendar(Calendar):

parser = ArgumentParser()
parser.add_argument('--calendar', type=FinalCalendar)
cfg = parser.parse_path('config.yaml')
  firstweekday: 1

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:

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

This yaml could be parsed as follows:

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

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

>>> parser = ArgumentParser(parser_mode='omegaconf')
>>> parser.add_argument('--server', type=ServerOptions)
>>> parser.add_argument('--client', type=ClientOptions)
>>> parser.add_argument('--config', action=ActionConfigFile)

>>> cfg = parser.parse_args(['--config=example.yaml'])
>>> cfg.client.url


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


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 .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_subcomm2 = ArgumentParser()
parser = ArgumentParser(prog='app')
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'])
Namespace(op0=None, subcomm1=Namespace(op1='val1'), subcommand='subcomm1')
>>> parser.parse_args(['--op0', 'val0', 'subcomm2', '--op2', 'val2'])
Namespace(op0='val0', subcomm2=Namespace(op2='val2'), subcommand='subcomm2')

Parsing config files with .ArgumentParser.parse_path or .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
  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))

>>> 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:

from jsonargparse import ArgumentParser, ActionConfigFile
parser = ArgumentParser(parser_mode='jsonnet')
parser.add_argument('--config', action=ActionConfigFile)
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:

from jsonargparse import ArgumentParser, ActionJsonnet, ActionJsonnetExtVars
parser = ArgumentParser()

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

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.

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) and the .ActionPathList class.

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
  info: data/info.db

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

from jsonargparse import ArgumentParser
from jsonargparse.typing import Path_fr
parser = ArgumentParser()
parser.add_argument('', 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 will be an instance of the .Path class that allows to get both the original relative path as included in the yaml file, or the corresponding absolute path:

>>> str(

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.

The content of a file that a .Path instance references can be read by using the .Path.get_content method. For the previous example would be info_db =

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 the .ActionPathList is used and as argument either the path to a file listing the paths is given or the special '-' string for reading the list from stdin. Example:

from jsonargparse import ActionPathList
parser.add_argument('--list', action=ActionPathList(mode='fr'))
cfg = parser.parse_args(['--list', 'paths.lst'])  # Text file with paths
cfg = parser.parse_args(['--list', '-'])          # List from stdin

If nargs='+' is given to add_argument with .ActionPathList then a single list is generated including all paths in all provided lists.


The .Path class is currently not fully supported in windows.

Parsing URLs

The .path_type function also supports URLs which after parsing, the .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 validators and requests python packages are 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 according to validators.url.url then 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 .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: .ArgumentParser.parse_path, .ArgumentParser.get_defaults (default_config_files argument), .ActionConfigFile, .ActionJsonSchema, .ActionJsonnet and .ActionParser. This means that a tool that can receive a configuration file via .ActionConfigFile is able to get the content from a URL, thus something like the following would work: --config

Boolean arguments

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:

>>> parser.add_argument('--op1', type=bool, default=False)
>>> parser.add_argument('--op2', type=bool, default=True)
>>> 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:

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'}.

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')
outer_parser = ArgumentParser(prog='app2')
    title='Inner node title',

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 .ActionConfigFile 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 .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 tool would be:

#!/usr/bin/env python3

from typing import Optional
from jsonargparse import ArgumentParser

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


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

$ chmod +x
$ eval "$(register-python-argcomplete"

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

Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

jsonargparse-4.10.0.tar.gz (171.5 kB view hashes)

Uploaded source

Built Distribution

jsonargparse-4.10.0-py3-none-any.whl (147.9 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page