Prerequisites

Your must install beartype in your Python environment in order to complete this How-To guide.

Add Enhanced Runtime Type-Checking to a Hydra App#

Many Python libraries (e.g. NumPy, PyTorch) include so-called type hints in the signatures of their functions and classes; a type-hint documents the type of values that an input parameter expects. Let’s turn these hints into promises! We can modify the configs in our Hydra app so that its configured values are verified against these types at runtime.

In this How-To guide we will:

Install a lightweight runtime type-checker, beartype.
Create “toy” library code that includes type annotations.
Design configs for a Hydra app that will leverage beartype’s type-checking throughout its interface.
Test that bad input values get “caught” by our type-checker.

Tip

hydra-zen also provides support for the popular runtime type-checker pydantic. It can be used instead of beartype. Consult the reference for validates_with_pydantic() to adapt this How-To accordingly.

To begin, we will install beartype in our Python environment.

1: Installing beartype.#

$ pip install beartype

Next, let’s create a toy example of some library code that uses type hints, but does not leverage runtime checking of it types. Create a script called toy_library.py, containing the following code.

2: Creating toy example of library code. Contents of toy_library.py.#

from typing import Union, Sequence, TypeAlias

from beartype.vale import Is
from typing_extensions import Annotated

PositiveInt: TypeAlias = Annotated[int, Is[lambda x: x >= 0]]

def process_age(age: PositiveInt):
    return age

def process_shape(shape: Union[int, Sequence[int]]):
    return shape

Note that the annotation PositiveInt indicates that an associated value should not only be an int, but also have a non-negative value. Whereas

Annotation of shape#

Union[int, Sequence[int]]

indicates that a value should either be an int or any sequence (list, tuple, etc.) of ints.

Supposing that our Hydra app will configure these two “toy” functions, let’s design their configs so that beartype will validate their configured values upon instantiation.

Open a Python console – or Jupyter notebook – in the same directory as toy_library.py and create the following configs.

3: Creating configs that include beartype validation.#

>>> from toy_library import process_age, process_shape

>>> from hydra_zen import make_custom_builds_fn
>>> from hydra_zen.third_party.beartype import validates_with_beartype

>>> builds = make_custom_builds_fn(
...     populate_full_signature=True,
...     zen_wrappers=validates_with_beartype,
...     hydra_convert="all",
... )

>>> ConfAge = builds(process_age)
>>> ConfShape = builds(process_shape)

Finally, let’s check that our configured values are validated as-expected. In the same console, verify that you can replicate the following behavior.

4: Test that the type-checker catches bad configured values.#

>>> from hydra_zen import instantiate

>>> instantiate(ConfAge, age=12)  # OK
12

>>> instantiate(ConfAge, age=-100)  # Bad: negative int
BeartypeCallHintPepParamException- process_age() parameter age=-100 violates type
hint [...]

>>> instantiate(ConfAge, age="twelve")  # Bad: not an int
BeartypeCallHintPepParamException- process_age() parameter age='twelve' violates
type hint [...]

>>> instantiate(ConfShape, shape=3)  # OK
3

>>> instantiate(ConfShape, shape=[1, 2, 5])  # OK
[1, 2, 5]

>>> instantiate(ConfShape, shape=["a", "b"])  # Bad: not a sequence of ints
BeartypeCallHintPepParamException- process_shape() parameter shape=['a', 'b']
violates type hint [...]

Awesome! Now mis-configured values will have a bear of a time getting past our app’s type-checked interface 🐻.

References