One of the big problems I have is cleanly converting from an internal datatype to an external datatype. We can all do it the not so clean way, however I think this add too much mess.

I can use libraries that read from a filetype to a Python object, however some libraries don't allow you to convert the data from one type to another. Most libraries also don't allow converting from one structure to another, so if you need data to be nested when it comes flat, you have to perform the conversion manually.

This library still uses these filetype libraries to convert to a Python object. It just offloads some of the work from these libraries. And adds some features I don't think will be added to these libraries.

This library consists of two public classes; Converter and Converters. These work almost completely independently. A short explanation of most of the code is:

  • Converters defines some property functions, these interface with Converter._obj to convert to and from the base class.
  • ron is used to raise when Converter._obj returns a 'null' (A BuilderObject), this is as we build a BuilderObject before building the actual class. This allows you to initialize using setattr, rather than just passing a dictionary. Which I find to be a little cleaner at times.
    This should be used whenever you get data from Converter._obj.
  • BuilderObject is a simple object that defaults nested objects to itself. This means we can build nested datatypes without having to build the objects themselves - as we don't have the data.
  • Converter is a small unobtrusive class to convert to and from the base class and itself. Providing T when using the class is required for the code to work.
from datetime import datetime
from typing import Generic, TypeVar, Type, get_type_hints, Union, List, Optional, Tuple, Any

__all__ = ['ron', 'Converter', 'Converters']

T = TypeVar('T')

class BuilderObject:
    def __init__(self):
        super().__setattr__('__values', {})

    def __getattr__(self, name):
        return super().__getattribute__('__values').setdefault(name, BuilderObject())

    def __setattr__(self, name, value):
        super().__getattribute__('__values')[name] = value

    def __delattr__(self, name):
        del super().__getattribute__('__values')[name]

def _build(base: Type[T], values: Union[BuilderObject, dict]) -> T:
    """Build the object recursively, utilizes the type hints to create the correct types"""
    types = get_type_hints(base)
    if isinstance(values, BuilderObject):
        values = super(BuilderObject, values).__getattribute__('__values')
    for name, value in values.items():
        if isinstance(value, BuilderObject) and name in types:
            values[name] = _build(types[name], value)
    return base(**values)

def _get_args(obj: object, orig: Type) -> Optional[Tuple[Type]]:
    """Get args from obj, filtering by orig type"""
    bases = getattr(type(obj), '__orig_bases__', [])
    for b in bases:
        if b.__origin__ is orig:
            return b.__args__
    return None

class Converter(Generic[T]):
    _obj: T

    def __init__(self, **kwargs) -> None:
        self._obj = BuilderObject()
        for name, value in kwargs.items():
            setattr(self, name, value)

    def build(self, exists_ok: bool=False) -> T:
        """Build base object"""
        t = _get_args(self, Converter)
        if t is None:
            raise ValueError('No base')
        base_cls = t[0]
        if isinstance(self._obj, base_cls):
            if not exists_ok:
                raise TypeError('Base type has been built already.')
            return self._obj
        self._obj = _build(base_cls, self._obj)
        return self._obj

    def from_(cls, b: T):
        """Build function from base object"""
        c = cls()
        c._obj = b
        return c

def ron(obj: T) -> T:
    """Error on null result"""
    if isinstance(obj, BuilderObject):
        raise AttributeError()
    return obj

TPath = Union[str, List[str]]

class Converters:
    def _read_path(path: TPath) -> List[str]:
        """Convert from public path formats to internal one"""
        if isinstance(path, list):
            return path
        return path.split('.')

    def _get(obj: Any, path: List[str]) -> Any:
        """Helper for nested `getattr`s"""
        for segment in path:
            obj = getattr(obj, segment)
        return obj

    def property(cls, path: TPath, *, get_fn=None, set_fn=None):
        Allows getting data to and from `path`.

        You can convert/type check the data using `get_fn` and `set_fn`. Both take and return one value.
        p = ['_obj'] + cls._read_path(path)

        def get(self):
            value = ron(cls._get(self, p))
            if get_fn is not None:
                return get_fn(value)
            return value

        def set(self, value: Any) -> Any:
            if set_fn is not None:
                value = set_fn(value)
            setattr(cls._get(self, p[:-1]), p[-1], value)

        def delete(self: Any) -> Any:
            delattr(cls._get(self, p[:-1]), p[-1])

        return property(get, set, delete)

    def date(cls, path: TPath, format: str):
        """Convert to and from the date format specified"""
        def get_fn(value: datetime) -> str:
            return value.strftime(format)

        def set_fn(value: str) -> datetime:
            return datetime.strptime(value, format)

        return cls.property(path, get_fn=get_fn, set_fn=set_fn)

An example of using this code is:

from dataclasses import dataclass
from datetime import datetime
from converters import Converter, Converters

from dataclasses_json import dataclass_json

class Range:
    start: datetime
    end: datetime

class Base:
    date: datetime
    range: Range

class International(Converter[Base]):
    date: str = Converters.date('date', '%d/%m/%y %H:%M')
    start: str = Converters.date('range.start', '%d/%m/%y %H:%M')
    end: str = Converters.date('range.end', '%d/%m/%y %H:%M')

class American(Converter[Base]):
    date: str = Converters.date('date', '%m/%d/%y %H:%M')
    start: str = Converters.date('range.start', '%m/%d/%y %H:%M')
    end: str = Converters.date('range.end', '%m/%d/%y %H:%M')

if __name__ == '__main__':
    i = International.from_json('''{
        "date": "14/02/19 12:00",
        "start": "14/02/19 12:00",
        "end": "14/02/19 12:00"
    b = i.build()
    a = American.from_(b)

    FORMAT = '{1}:\n\tdate: {0.date}\n\tstart: {0.range.start}\n\tend: {0.range.end}'
    FORMAT_C = '{1}:\n\tdate: {0.date}\n\tstart: {0.start}\n\tend: {0.end}'
    print(FORMAT.format(b, 'b'))
    print(FORMAT_C.format(a, 'a'))
    print(FORMAT_C.format(i, 'i'))
    print('\nupdate b.date')
    b.date = datetime(2019, 2, 14, 12, 30)
    print(FORMAT.format(b, 'b'))
    print(FORMAT_C.format(a, 'a'))
    print(FORMAT_C.format(i, 'i'))
    print('\nupdate b.range.start')
    b.range.start = datetime(2019, 2, 14, 13, 00)
    print(FORMAT.format(b, 'b'))
    print(FORMAT_C.format(a, 'a'))
    print(FORMAT_C.format(i, 'i'))

    print('\njson dump')

From a code review I mostly want to focus on increasing the readability of the code. I also want to keep Converter to contain all of the logic, whilst also being very transparent so that most libraries like dataclasses_json work with it. I don't care about performance just yet.

  • 2
    \$\begingroup\$ @BenoîtPilatte I don't think discussing this in the comments will achieve any good. Please feel free to post an answer if you wish. An explanation on my abuse of the 'hinting' system, an explanation of where this 'call stack hell' is and an explanation of why I want code to convert datatypes (the opposite of what I want) would be good to see in this answer. \$\endgroup\$
    – Peilonrayz
    Commented Feb 18, 2019 at 14:28
  • \$\begingroup\$ Say Range has a third attribute step: int, how would you refer to it in your external classes? step: int = Converters.property('range.step')? \$\endgroup\$ Commented Feb 26, 2019 at 22:23
  • \$\begingroup\$ @MathiasEttinger Yeah assuming it's defined in Range too. If you need to convert to int from str then you can use step: int = Converters.property('range.step', get_fn=str, set_fn=int). \$\endgroup\$
    – Peilonrayz
    Commented Feb 26, 2019 at 22:27

2 Answers 2


Let's get from the simplest thing to change to the hardest one:


This is actually a recursive dictionary, which is more easily achieved using either collections.defaultdict or by implementing dict.__missing__. Since you also want __getattr__ to act like __getitem__ and so on, I’d go the dict subclass route:

class BuilderObject(dict):
    def __missing__(self, item):
        self[item] = missing = BuilderObject()
        return missing

    def __getattr__(self, item):
        return self[item]

    def __setattr__(self, item, value):
        self[item] = value

    def __delattr__(self, item):
        del self[item]

Simpler to read and understand. I’m not fond of the name however, but couldn't come up with something better.


These are in fact descriptors in disguise. Instead of writing the getter, setter and deleter functions to feed to property, you could as well write them as __get__, __set__ and __delete__ method on a custom class. Also operator.attrgetter is your friend, no need to rewrite it yourself:

class AttributeProxy:
    def __init__(self, path: str, *, get_fn=None, set_fn=None):
        self.__path = '_obj.' + path
        self.__parent, self.__attribute_name = self.__path.rsplit('.', 1)
        self.__getter = get_fn
        self.__setter = set_fn

    def __get__(self, instance, owner):
        if not issubclass(owner, Converter):
            raise RuntimeError('cannot use Property descriptors on non Converter types')
        if instance is None:
            return self
        value = operator.attrgetter(self.__path)(instance)
        if isinstance(value, BuilderObject):
            raise AttributeError
        if self.__getter is not None:
            value = self.__getter(value)
        return value

    def __set__(self, instance, value):
        if self.__setter is not None:
            value = self.__setter(value)
        setattr(operator.attrgetter(self.__parent)(instance), self.__attribute_name, value)

    def __delete__(self, instance):
        delattr(operator.attrgetter(self.__parent)(instance), self.__attribute_name)

class DateProxy(AttributeProxy):
    def __init__(self, path, format):
            get_fn=lambda value: value.strftime(format),
            set_fn=lambda value: datetime.strptime(value, format)


The main purpose of this function is to check that, when one derives from Converter, they provide a specialization of T; and optionally retrieve that specialization. I find the implementation somewhat cryptic and potentially missing some edge cases. I fiddled with a metaclass so the check is performed very early in the program. This is the main advantage of the approach as you would get an error message right away, not latter on in an unrelated section of the code:

class CheckBaseExist(type):
    def __new__(mcls, name, bases, attrs):
        cls = super().__new__(mcls, name, bases, attrs)
        if not issubclass(cls, Generic):
            raise TypeError('CheckBaseExist metaclass should be used on typing.Generic subclasses')

        if Generic not in bases:
            # We already know that klass is a subclass of Generic so it must define __orig_bases__
                base, = cls.__orig_bases__
            except ValueError:
                raise TypeError('cannot use more than one specialization of a base CheckBaseExist class in the inheritance tree') from None
            if base.__origin__ is Generic:
                raise TypeError('no specialization provided when inheriting from a base CheckBaseExist class')
            generic_subclasses = ' or '.join(
                for klass in bases
                if klass is not Generic and issubclass(klass, Generic)
            if generic_subclasses:
                raise TypeError(f'cannot use typing.Generic as a common base class with {generic_subclasses}')
        return cls

Not too fond of the name either… Usage being:

class Converter(Generic[T], metaclass=CheckBaseExist):
    _obj: T

    def __init__(self, **kwargs) -> None:
        self._obj = BuilderObject()
        for name, value in kwargs.items():
            setattr(self, name, value)

    def to_base(self, exists_ok: bool=False) -> T:
        """Build base object"""
        base_cls = self.__class__.__orig_bases__[0].__args__[0]
        if isinstance(self._obj, BuilderObject):
            self._obj = _build(base_cls, self._obj)
        elif not exists_ok:
            raise RuntimeError('Base type has been built already.')
        return self._obj

    def from_base(cls, base: T):
        """Build function from base object"""
        instance = cls()
        instance._obj = base
        return instance

This works as:

  1. we checked in the metaclass that self.__class__.__orig_bases__ contains a single item;
  2. Generic[T] ensures that __args__ contains a single item.


I really dislike the fact that this function relies on typing.get_type_hints but couldn't come up with something clean that doesn't use it, or at least does it optionally. Maybe an extra argument in the AttributeProxy constructor, defaulting to None. I’m not a huge fan of it, but you’ll have to provide type hints somehow anyway.

This is important in case you want to convert to/from objects in external libraries that don't use those type hints, so you must implement a fallback mechanism.

  • \$\begingroup\$ Holy moley this is a sick answer. Thank you for staying up so late to write it! I'm going to need to digest it again tomorrow too to fully understand it. Whilst your pros for the meta class are clear, I'm not sure they outweigh the cons. 1. It'll make converting to Python <3.7 hard, as Generic has its own (iirc private) metaclass, and it'll mean it won't work with 'filetype libraries' that require metaclasses. I think I'll fix this by adding another class variable to hold the base, and have CheckBaseExist automagically set it. Allowing Converter and GenericConverter classes. (Thoughts?) \$\endgroup\$
    – Peilonrayz
    Commented Feb 27, 2019 at 2:47
  • \$\begingroup\$ I'm not sure "This is important in case you want to convert to/from objects in external libraries that don't use those type hints" is quite true. This is as it's passed Base not a Converter, and so it should be separated from these external libraries. Or have I misunderstood you here? \$\endgroup\$
    – Peilonrayz
    Commented Feb 27, 2019 at 2:50
  • \$\begingroup\$ @Peilonrayz No, it means that you can't do Converter[socket.socket] for instance, or anything that doesn't use type hints. \$\endgroup\$ Commented Feb 27, 2019 at 6:43

There are a couple of things that come to mind when reading your code. I'll write them down, in no particular order of importance.

Your example imports your library like this:

from converters import Converter, Converters

That's a major code smell. You import both a Converter and Converters from a file named converters. If you have 3 things with the same name, you should name them better. Why do you have a Converter and a Converters class anyway? I'd expect one to be a collection of the other, but since Converter already takes a template and is generic, what the heck do I need it's multiple for? It's not intuitive and probably violates the Zen of Python on half a dozen principles.

I see a lot of single-letter variables. While T is somewhat acceptable here, the rest is not. i = International.from_json( What? No. i is an index or some other integer, not something much more complicated than that.

b = i.build()
a = American.from_(b)

Please, no. American and International are terrible names for classes anyway. You could use it as a sub-class or sub-type or something, or an instance of a class if the class makes clear it's a date of some sort, but don't make an American class.

Now we're talking about those classes anyway, did you notice the absurd amount of repetition?

class International(Converter[Base]):
    date: str = Converters.date('date', '%d/%m/%y %H:%M')
    start: str = Converters.date('range.start', '%d/%m/%y %H:%M')
    end: str = Converters.date('range.end', '%d/%m/%y %H:%M')

class American(Converter[Base]):
    date: str = Converters.date('date', datetime_format)
    start: str = Converters.date('range.start', datetime_format)
    end: str = Converters.date('range.end', datetime_format)

So, a class has 3 lines and all of those lines contain either '%d/%m/%y %H:%M' or '%m/%d/%y %H:%M'. Have you considered making something like this instead?

class TerriblyNamedGeneric(Converter[Base], datetime_format):
    date: str = Converters.date('date', datetime_format)
    start: str = Converters.date('range.start', datetime_format)
    end: str = Converters.date('range.end', datetime_format)

That's still not pretty and it can probably be done with even less repetition, but you get the idea.

The rest of your code is riddled with ambiguity.

def from_(cls, b: T):
    """Build function from base object"""
    c = cls()
    c._obj = b
    return c

Why is a build function named from_? What is a cls (if it's not short for one of these, pick a better name) and why is that entire function just a jumble of one-letter variable names?

You say you want a review focussed on the readability. In short, I don't think it's that readable. It may work like a charm, but the readability leaves much to be desired.

  • 1
    \$\begingroup\$ Good answer, as discussed in chat I think ObjectConverter and PropertyConverters would be better names. I think from_ is standard, and cls is standard in classmethods. But the rest about from_ I agree with. Thanks :) \$\endgroup\$
    – Peilonrayz
    Commented Feb 26, 2019 at 15:38
  • \$\begingroup\$ Why would from converters import Converter, Converters be inherently wrong? I often do from datetime import datetime or from pprint import pprint or import socket; sock = socket.socket(…) \$\endgroup\$ Commented Feb 26, 2019 at 22:10
  • \$\begingroup\$ @MathiasEttinger We talked about that in chat. It's not so much wrong, as in it's unclear why you'd need to import both Converter and Converters from a library called converters. It's a naming problem. The whole thing leads to ambiguity. Plenty of libraries are plagued with it, I know. That doesn't make it right. \$\endgroup\$
    – Mast
    Commented Feb 27, 2019 at 14:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.