pytypes is a typing toolbox w.r.t. PEP 484 (PEP 526 on the road map, later also 544 if it gets accepted).

Its main features are currently

• @typechecked decorator for runtime typechecking with support for stubfiles and type comments
• @override decorator that asserts existence of a type-compatible parent method
• @annotations decorator to turn type info from stubfiles or from type comments into __annotations__
• @typelogged decorator observes function and method calls at runtime and generates stubfiles from acquired type info
• service functions to apply these decorators module wide or even globally, i.e. runtime wide
• typechecking can alternatively be done in decorator-free manner (friendlier for debuggers)
• all the above decorators work smoothly with OOP, i.e. with methods, static methods, class methods and properties, even if classes are nested
• converter for stubfiles to Python 2.7 compliant form
• lots of utility functions regarding types, e.g. a Python 2.7 compliant and actually functional implementation of get_type_hints
• full Python 2.7 support for all these features

An additional future goal will be integration with the Java typing system when running on Jython. Along with this, some generator utilities to produce type-safe Java bindings for Python frameworks are planned.

In wider sense, PEP 484-style type annotations can be used to build type safe interfaces to allow also other programming languages to call into Python code (kind of reverse FFI). In this sense the project name refers to 'ctypes', which provides Python-bindings of C.

All described features of pytypes were carefully implemented such that they are equally workable on CPython 3.5, 3.6, 2.7 and on Jython 2.7.1 (other interpreters might work as well, but were not yet tested). For Python 2.7, pytypes fully supports type-annotations via type comments. It also supports Python 2.7-style type annotations in Python 3.5-code to allow easier 2.7/3.5 multi-version development. Python 3.7 and 3.8 are mostly supported, but some bugs still need to be fixed.

There have been earlier approaches for runtime-typechecking. However, most of them predate PEP 484 or lack some crucial features like support of Python 2.7 or support of stubfiles. Also, none of them features a typechecking override decorator. There were separate approaches for override decorators, but these usually don't consider PEP 484 at all. So we decided that it's time for a new runtime typechecking framework, designed to support PEP 484 from the roots, including its extensive features like (Python 2.7-style-)type comments and stub files.

pytypes provides a rich set of utilities for runtime typechecking.

Decorator applicable to functions, methods, properties and classes. Asserts compatibility of runtime argument and return values of all targeted functions and methods w.r.t. PEP 484-style type annotations of these functions and methods. This supports stubfiles and type comments and is thus workable on Python 2.7.

Running Python with the '-O' flag, which also disables assert statements, turns off typechecking completely. Alternatively, one can modify the flag pytypes.checking_enabled.

Note that this must be done right after import of pytypes, because it affects the way how @typechecked decorator works. For modules that were imported with this flag disabled, typechecking cannot be turned on later on within the same runtime.

from pytypes import typechecked

@typechecked
def some_function(a, b, c):
    # type: (int, str, List[Union[str, float]]) -> int
    return a+len(b)+len(c)

from pytypes import typechecked

@typechecked
def some_function(a: int, b: str, c: List[Union[str, float]]) -> int:
    return a+len(b)+len(c)

The decorators in this section allow type-safe method overriding.

Decorator applicable to methods only. For a version applicable also to classes or modules use auto_override. Asserts that for the decorated method a parent method exists in its mro. If both the decorated method and its parent method are type annotated, the decorator additionally asserts compatibility of the annotated types. Note that the return type is checked in contravariant manner. A successful check guarantees that the child method can always be used in places that support the parent method's signature. Use pytypes.check_override_at_runtime and pytypes.check_override_at_class_definition_time to control whether checks happen at class definition time or at "actual runtime".

The following rules apply for override checking:

• a parent method must exist
• the parent method must have call-compatible signature (e.g. same number of args)
• arg types of parent method must be more or equal specific than arg types of child
• return type behaves contravariant - parent method must have less or equal specific return type than child

from pytypes import override

class some_baseclass():
    def some_method1(self, a: int) -> None: ...
    def some_method2(self, a: int) -> None: ...
    def some_method3(self, a: int) -> None: ...
    def some_method4(self) -> int: ...

class some_subclass(some_baseclass):
    @override
    def some_method1(self, a: float) -> None: ...

    @override
    def some_method2(self, a: str) -> None: ...

    @override
    def some_metd3(self, a: int) -> None: ...

    @override
    def some_method4(self) -> float: ...

• some_method1: override check passes
• some_method2: override check fails because type is not compatible
• some_method3: override check fails because of typo in method name
• some_method4: override check fails because return type must be more or equal specific than parent

Decorator applicable to methods and classes. Works like override decorator on type annotated methods that actually have a type annotated parent method. Has no effect on methods that do not override anything. In contrast to plain override decorator, auto_override can be applied easily on every method in a class or module. In contrast to explicit override decorator, auto_override is not suitable to detect typos in spelling of a child method's name. It is only useful to assert compatibility of type information (note that return type is contravariant). Use pytypes.check_override_at_runtime and pytypes.check_override_at_class_definition_time to control whether checks happen at class definition time or at "actual runtime".

The following rules apply, if a parent method exists:

• the parent method must have call-compatible signature (e.g. same number of args)
• arg types of parent method must be more or equal specific than arg types of child
• return type behaves contravariant - parent method must have less or equal specific return type than child

Compared to ordinary override decorator, the rule “a parent method must exist” is not applied here. If no parent method exists, auto_override silently passes.

Decorator applicable to functions, methods, properties and classes. Methods with type comment will have type hints parsed from that string and get them attached as __annotations__ attribute. Methods with either a type comment or ordinary type annotations in a stubfile will get that information attached as __annotations__ attribute (also a relevant use case in Python 3). Behavior in case of collision with previously (manually) attached __annotations__ can be controlled using the flags pytypes.annotations_override_typestring and pytypes.annotations_from_typestring.

Decorator applicable to functions, methods, properties and classes. It observes function and method calls at runtime and can generate stubfiles from acquired type info.

One can disable typelogging via the flag pytypes.typelogging_enabled.

Note that this must be done right after import of pytypes, because it affects the way how @typelogged decorator works. For modules that were imported with this flag disabled, typelogging cannot be turned on later on within the same runtime.

Assume you run a file ./script.py like this:

from pytypes import typelogged

@typelogged
def logtest(a, b, c=7, *var, **kw): return 7, a, b

@typelogged
class logtest_class(object):
    def logmeth(self, b): return 2*b

    @classmethod
    def logmeth_cls(cls, c): return len(c)

    @staticmethod
    def logmeth_static(c): return len(c)

    @property
    def log_prop(self): return self._log_prop

    @log_prop.setter
    def log_prop(self, val): self._log_prop = val

logtest(3, 2, 5, 6, 7, 3.1, y=3.2, x=9)
logtest(3.5, 7.3, 5, 6, 7, 3.1, y=3.2, x=9)
logtest('abc', 7.3, 5, 6, 7, 3.1, y=2, x=9)
lcs = logtest_class()
lcs.log_prop = (7.8, 'log')
lcs.log_prop
logtest_class.logmeth_cls('hijk')
logtest_class.logmeth_static(range(3))

pytypes.dump_cache()

Alternatively you can use the TypeLogger profiler:

from pytypes import TypeLogger

def logtest(a, b, c=7, *var, **kw): return 7, a, b

class logtest_class(object):
    def logmeth(self, b): return 2*b

    @classmethod
    def logmeth_cls(cls, c): return len(c)

    @staticmethod
    def logmeth_static(c): return len(c)

    @property
    def log_prop(self): return self._log_prop

    @log_prop.setter
    def log_prop(self, val): self._log_prop = val

with TypeLogger():
    logtest(3, 2, 5, 6, 7, 3.1, y=3.2, x=9)
    logtest(3.5, 7.3, 5, 6, 7, 3.1, y=3.2, x=9)
    logtest('abc', 7.3, 5, 6, 7, 3.1, y=2, x=9)
    lcs = logtest_class()
    lcs.log_prop = (7.8, 'log')
    lcs.log_prop
    logtest_class.logmeth_cls('hijk')
    logtest_class.logmeth_static(range(3))

Note that this will produce more stubs, i.e. also for indirectly used modules, because the profiler will handle every function call. To scope a specific module at a time use pytypes.typelogged on that module or its name. This should be called on a module after it is fully loaded. To use it inside the scoped module (e.g. for __main__) apply it right after all classes and functions are defined.

Any of the examples above will create the following file in ./typelogger_output:

script.pyi:

from typing import Tuple, Union

def logtest(a: Union[float, str], b: float, c: int, *var: float, **kw: Union[float, int]) -> Union[Tuple[int, float, float], Tuple[int, str, float]]: ...

class logtest_class(object):
    def logmeth(self, b: int) -> int: ...

    @classmethod
    def logmeth_cls(cls, c: str) -> int: ...

    @staticmethod
    def logmeth_static(c: range) -> int: ...

    @property
    def log_prop(self) -> Tuple[float, str]: ...

    @log_prop.setter
    def log_prop(self, val: Tuple[float, str]) -> None: ...

Use pytypes.dump_cache(python2=True) to produce a Python 2.7 compliant stubfile.

By default, pytypes performs pytypes.dump_cache() at exit, i.e. writes typelog as a Python 3 style stubfile. Use pytypes.dump_typelog_at_exit to control this behavior. Use pytypes.dump_typelog_at_exit_python2 to write typelog as a Python 2 style stubfile.

Note that global mode is experimental.

The pytypes decorators @typechecked, @auto_override, @annotations and @typelogged can be applied module wide by explicitly calling them on a module object or a module name contained in sys.modules. In such a case, the decorator is applied to all functions and classes in that module and recursively to all methods, properties and inner classes too.

Warning: If A decorator is applied to a partly imported module, only functions and classes that were already defined are affected. After the module imported completely, the decorator is applied to the remaining functions and classes. In the meantime, internal code of that module can circumvent the decorator, e.g. can make module-internal calls that are not typechecked.

The pytypes decorators @typechecked and @typelogged have corresponding profiler implementations TypeChecker and TypeLogger. You can conveniently install them globally via enable_global_typechecked_profiler() and enable_global_typelogged_profiler().

Alternatively you can apply them in a with-context:

from pytypes import TypeChecker

def agnt_test(v):
    # type: (str) -> int
    return 67

with TypeChecker():
    agnt_test(12)

One glitch is to consider in case you want to catch TypeCheckError (i.e. ReturnTypeError or InputTypeError as well) and continue execution afterwards. The TypeChecker would be suspended unless you call restore_profiler, e.g.:

from pytypes import TypeChecker, restore_profiler

def agnt_test(v):
    # type: (str) -> int
    return 67

with TypeChecker():
    try:
        agnt_test(12)
    except TypeCheckError:
        restore_profiler()
        # handle error....

Note that the call to restore_profiler must be performed by the thread that raised the error.

Alternatively you can enable pytypes.warning_mode = True to raise warnings rather than errors. (This only helps if you don't use filterwarnings("error") or likewise.)

The pytypes decorators @typechecked, @auto_override, @annotations and @typelogged can be applied globally to all loaded modules and subsequently loaded modules. Modules that were loaded while typechecking or typelogging was disabled will not be affected. Apart from that this will affect every module in the way described above. Note that we recommend to use the profilers explained in the previous section if global typechecking or typelogging is required. Use this feature with care as it is still experimental and can notably slow down your python runtime. In any case, it is intended for debugging and testing phase only.

• To apply @typechecked globally, use pytypes.set_global_typechecked_decorator
• To apply @auto_override globally, use pytypes.set_global_auto_override_decorator
• To apply @annotations globally, use pytypes.set_global_annotations_decorator
• To apply @typelogged globally, use pytypes.set_global_typelogged_decorator

Warning: If the module that performs the ``pytypes.set_global_xy_decorator``-call is not yet fully imported, the warning regarding module-wide decorators (see above) applies to that module in the same sense. I.e. functions and classes that were not yet defined, will be covered only once the module-import has fully completed.

All the above decorators work smoothly with OOP. You can safely apply @typechecked, @annotations and @typelogged on methods, abstract methods, static methods, class methods and properties. @override is – already by semantics – only applicable to methods, @auto_override is additionally applicable to classes and modules.

pytypes also takes care of inner classes and resolves name space properly. Make sure to apply decorators from pytypes on top of @staticmethod, @classmethod, @property or @abstractmethod rather than the other way round. This is because OOP support involves some special treatment internally, so OOP decorators must be visible to pytypes decorators. This also applies to old-style classes.

For now, @override cannot be applied to __init__, because __init__ typically extends the list of initialization parameters and usually uses super to explicitly serve a parent's signature. The purpose of @override is to avoid typos and to guarantee that the child method can always be used as a fill in for the parent in terms of signature and type information. Both aspects are hardly relevant for __init__:

• a typo is unlikely and would show up quickly for various reasons
• when creating an instance the caller usually knows the exact class to instantiate and thus its signature

For special cases where this might be relevant, @typechecked can be used to catch most errors.

Utility functions described in this section can be directly imported from the pytypes module. Only the most important utility functions are listed here.

Resembles typing.get_type_hints, but is also workable on Python 2.7 and searches stubfiles for type information. Also on Python 3, this takes type comments into account if present.

Works like get_type_hints, but returns types as a sequence rather than a dictionary. Types are returned in declaration order of the corresponding arguments.

This function mimics typeguard syntax and semantics. It can be applied within a function or method to check argument values to comply with type annotations. It behaves similar to @typechecked except that it is not a decorator and does not check the return type. A decorator less way for argument checking yields less interference with some debuggers.

This function works like check_argument_types, but applies to the return value. Because it is impossible for pytypes to automatically figure out the value to be returned in a function, it must be explicitly provided as the value-parameter.

Works like isinstance, but supports PEP 484 style types from typing module.

If cls contains unbound TypeVar s and bound_Generic is provided, this function attempts to retrieve corresponding values for the unbound TypeVar s from bound_Generic.

Works like issubclass, but supports PEP 484 style types from typing module.

If subclass or superclass contains unbound TypeVar s and bound_Generic is provided, this function attempts to retrieve corresponding values for the unbound TypeVar s from bound_Generic.

Tries to construct a type for a given value. In contrast to type(...), deep_type does its best to fit structured types from typing as close as possible to the given value. E.g. deep_type((1, 2, 'a')) will return Tuple[int, int, str] rather than just tuple. Supports various types from typing, but not yet all. Also detects nesting up to given depth (uses pytypes.default_typecheck_depth if no value is given). If a value for max_sample is given, this number of elements is probed from lists, sets and dictionaries to determine the element type. By default, all elements are probed. If there are fewer elements than max_sample, all existing elements are probed.

Generates a nicely readable string representation of the given type. The returned representation is workable as a source code string and would reconstruct the given type if handed to eval, provided that globals/locals are configured appropriately (e.g. assumes that various types from typing have been imported). Used as type-formatting backend of ptypes' code generator abilities in modules typelogger and stubfile_2_converter. If tp contains unbound TypeVar s and bound_Generic is provided, this function attempts to retrieve corresponding values for the unbound TypeVar s from bound_Generic.

Writes cached observations by @typelogged into stubfiles.

Files will be created in the directory provided as 'path'; overwrites existing files without notice. Uses 'pyi2' suffix if 'python2' flag is given else 'pyi'. Resulting files will be Python 2.7 compliant accordingly.

Retrieves the item type from a PEP 484 generic or subclass of such. sq must be a typing.Tuple or (subclass of) typing.Iterable or typing.Container. Consequently this also works with typing.List, typing.Set and typing.Dict. Note that for typing.Dict and mapping types in general, the key type is regarded as item type. For typing.Tuple all contained types are returned as a typing.Union. If simplify == True some effort is taken to eliminate redundancies in such a union.

Retrieves the key and value types from a PEP 484 mapping or subclass of such. mp must be a (subclass of) typing.Mapping.

Retrieves the parameter value of a given TypeVar from a Generic. Returns None if the generic does not contain an appropriate value. Note that the TypeVar is compared by instance and not by name. E.g. using a local TypeVar T would yield different results than using typing.T despite the equal name.

Resolves forward references in in_type.

globs should be a dictionary containing values for the names that must be resolved in in_type. If globs is not provided, it will be created by __globals__ from the module named module_name, plus __locals__ from the last search_stack_depth stack frames, beginning at the calling function. This is to resolve cases where in_type and/or types it fw-references are defined inside a function.

To prevent walking the stack, set search_stack_depth=0. Ideally provide a proper globs for best efficiency. See util.get_function_perspective_globals for obtaining a globs that can be cached. util.get_function_perspective_globals works like described above.

Robust way to access obj.__orig_class__. Compared to a direct access this has the following advantages:

1. It works around python/typing#658.

2. It prevents infinite recursion when wrapping a method (obj is self or cls) and either

   • the object's class defines __getattribute__ or
   • the object has no __orig_class__ attribute and the object's class defines __getattr__.

   See discussion at pull request 53.

If default_to__class__ is True it returns obj.__class__ as final fallback. Otherwise, AttributeError is raised in failure case (default behavior).

Currently pytypes uses the python runtime, i.e. import, eval, dir and inspect to parse stubfiles and type comments. A runtime independent parser for stubfiles is a desired future feature, but is not yet available. This means that conventional PEP 484 stubfiles would not work on Python 2.7. To resolve this gap, pytypes features a converter script that can convert conventional stubfiles into Python 2.7 compliant form. More specifically it converts parameter annotations into type comments and converts ... syntax into pass.

As of this writing it does not yet support stubfiles containing the @overload decorator. Also, it does not yet convert type annotations of attributes and variables.

pytypes uses the suffix 'pyi2' for Python 2.7 compliant stubfiles, but does not require it. Plain 'pyi' is also an acceptable suffix (as far as pytypes is concerned), because Python 2.7 compliant stubfiles can also be used in Python 3.

The main purpose of 'pyi2' suffix is to avoid name conflicts when conventional stubfiles and Python 2.7 compliant stubfiles coexist for the same module. In that case the pyi2 file will override the pyi file when running on Python 2.7.

Run stubfile_2_converter.py to leverage pytypes' stubfile converter capabilities:

python3 -m pytypes.stubfile_2_converter [options/flags] [in_file]

Use python3 -m pytypes.stubfile_2_converter -h to see detailed usage.

By default the out file will be created in the same folder as the in file, but with 'pyi2' suffix.

• support PEP 526
• support overloading
• support named tuple
• support async-related constructs from typing
• support notation for Positional-only arguments
• runtime independent parser for stubfiles

pytypes was created in 2016/17 by Stefan Richthofer.

• Alex Grönholm
• Mitar
• Ilya Kulakov
• Jonas
• MinJune Kim
• Uryga
• Steve Jessop
• David Barnett

pytypes is released under Apache 2.0 license. A copy is provided in the file LICENSE.

[email protected]