This is a PEP-561-compliant stub-only package which provides type information for matplotlib, numpy and pandas. The mypy type checker (or pytype or PyCharm) can recognize the types in these packages by installing this package.

Lots of functions are already typed, but a lot is still missing (numpy and pandas are huge libraries). Chances are you will see a message from Mypy claiming that a function does not exist when it actually does exist. If you encounter missing functions, we would be very happy for you to send a PR. If you are unsure of how to type a function, we can discuss it.

You can get this package from Pypi:

pip install data-science-types

To get the most up-to-date version, install it directly from GitHub:

pip install git+https://github.com/predictive-analytics-lab/data-science-types

Or clone the repository somewhere and do pip install -e ..

These are the kinds of things that can be checked:

import numpy as np

arr1: np.ndarray[np.int64] = np.array([3, 7, 39, -3])  # OK
arr2: np.ndarray[np.int32] = np.array([3, 7, 39, -3])  # Type error
arr3: np.ndarray[np.int32] = np.array([3, 7, 39, -3], dtype=np.int32)  # OK
arr4: np.ndarray[float] = np.array([3, 7, 39, -3], dtype=float)  # Type error: the type of ndarray can not be just "float"
arr5: np.ndarray[np.float64] = np.array([3, 7, 39, -3], dtype=float)  # OK

import numpy as np

arr1: np.ndarray[np.int64] = np.array([3, 7, 39, -3])
arr2: np.ndarray[np.int64] = np.array([4, 12, 9, -1])

result1: np.ndarray[np.int64] = np.divide(arr1, arr2)  # Type error
result2: np.ndarray[np.float64] = np.divide(arr1, arr2)  # OK

compare: np.ndarray[np.bool_] = (arr1 == arr2)

import numpy as np

arr: np.ndarray[np.float64] = np.array([[1.3, 0.7], [-43.0, 5.6]])

sum1: int = np.sum(arr)  # Type error
sum2: np.float64 = np.sum(arr)  # OK
sum3: float = np.sum(arr)  # Also OK: np.float64 is a subclass of float
sum4: np.ndarray[np.float64] = np.sum(arr, axis=0)  # OK

# the same works with np.max, np.min and np.prod

The goal is not to recreate the APIs exactly. The main goal is to have useful checks on our code. Often the actual APIs in the libraries is more permissive than the type signatures in our stubs; but this is (usually) a feature and not a bug.

We always welcome contributions. All pull requests are subject to CI checks. We check for compliance with Mypy and that the file formatting conforms to our Black specification.

You can install these dev dependencies via

pip install -e '.[dev]'

This will also install numpy, pandas and matplotlib to be able to run the tests.

We include a script that runs the CI checks that will be run when a PR is opened. To test these out locally, you need to install the type stubs in your environment. Typically, you would do this with

pip install -e .

Then use the check_all.sh script to run all tests:

./check_all.sh

Below we describe how to run the various checks individually, but check_all.sh should be easier to use.

The settings for Mypy are specified in the mypy.ini file in the repository. Just running

mypy tests

from the base directory should take these settings into account. We enforce 0 mypy errors.

We use Black to format the stub files. First install black and then run

black .

from the base directory.

python -m pytest -vv tests/

flake8 *-stubs

Apache 2.0