auto_serialize

This is a simple two-file serialization/deserialization library.

This is not a data validation library. For data validation, use pydantic. If you have a polyglot codebase, use pydantic. You won't care about the schema, but your colleagues will.

Installation

Install with pip install auto_serialize

Usage

Just inherit from AutoSerialize!

from auto_serialize import AutoSerialize

class MyClass(AutoSerialize):
    def __init__(self, a: int, b: tuple[dict[str, list[float]], str]):
        self._a = a
        self.b = b

test = MyClass(2, ({"hello": [4, 2]}, "world"))

# Go to and from json.
json_str = test.to_json()
test_again = MyClass.from_json(json_str)

json_str == '{"a":2,"b":[{"hello":[4,2]},"world"]}' # True
test_again == test # True.

# We can also serialize to yaml. More formats coming soon :)
test.to_yaml() == 'a: 2\nb: !!python/tuple\n- hello:\n  - 4\n  - 2\n- world\n' # True

AutoSerialize doesn't just support inbuilt types. It also supports unions, numpy arrays, heavily nested data types, and serialization/deserialization callbacks.

class OtherDemoClass(AutoSerialize):
    def __init__(self, value: int) -> None:
        self._value = value


class DemoClass(AutoSerialize):
    def __init__(
        self,
        a: int,
        b: list[float] | float,
        c: dict[str, OtherDemoClass | list[int]],
        arr: np.ndarray,
        maybe: str | None = None,
    ) -> None:
        self._a = a
        self._b = b
        self.c = c
        self.arr = arr
        self._maybe = maybe

    def before_serialize(self) -> None:
        self._a += 1

    def after_deserialize(self) -> None:
        if self._maybe is None:
            self._maybe = "world!"

other_demo_class = OtherDemoClass(5)
demo = DemoClass(
    1, 3.141, {"hello, ": other_demo_class}, np.array([4, 5, 6])
)

json_str = demo.to_json()
demo_again = DemoClass.from_json(json_str)

# The following assertions all pass.
assert demo_again._a == 2
assert demo_again._b == 3.141
assert demo_again.c == {"hello, ": other_demo_class}
assert np.array_equal(demo_again.arr, np.array([4, 5, 6]))
assert isinstance(demo_again.arr, np.ndarray)
assert demo_again._maybe == "world!"

Some other handy things you get from using AutoSerialize:

• A deepcopy() method
  • Currently implemented by serializing to json, then deserializing back.
• __eq__, __neq__ and __hash__ default implementations.
  • Currently implemented by mapping your AutoSerialize objects to python dictionaries.

Why?

This is a self serving repository.

I need to serialize a lot of stuff, and I have had this need for a long time. I have no time for 3rd party libraries that don't support my data format. If I need to serialize a numpy array with NaNs, I need to be able to implement that functionality in a very small amount of time.

I have posted this online in the hope that the internet will bully me into supporting more data formats, to build something really meaningful.

Implementation details; gotchas

AutoSerialize works by inspecting the type hints on __init__. If you don't use type hints, it won't work.

Union resolution is optimistic, going from left to right, and not all generics are checked for. As a result, if the above DemoClass was type hinted c: dict[str, list[int] | OtherDemoClass], The OtherDemoClass would actually get deserialized as a list! This is because the int generic isn't checked for, and OtherDemoClass is serialized to a dict, which is iterable! 😱

But, because a list of integers could never be deserialized as OtherDemoClass, using the ordering c: dict[str, OtherDemoClass | list[int]], will always work. If c was a list of integers, AutoSerialize would start by trying to deserialize the list of integers as an instance of OtherDemoClass. This would fail, and then AutoSerialize would fallback to the list[int] option, which would work! 🎉

Contributing

If you have an example class that won't serialize, submit an issue. I'll fix it and add your class to the test suite! Of course, feel free to submit a PR if you're up to fixing it yourself. 💪

Roadmap

Formats

I have an interest in hdf5 and protobuf formats, so it's likely that they will both end up being supported in the near future. If you have a particular need for either of these formats, please submit an issue. That will be enough motivation for me to get it done in no time.

Schema

It would be pretty easy to make this library have the same validation functionality that pydantic has. I haven't pushed myself to implement this because pydantic already exists. If there was ever a need, it would be done.