Coding Conventions

Code Formatting

We use yapf to format all of the code in DeepChem. Although it sometimes produces slightly awkward formatting, it does have two major benefits. First, it ensures complete consistency throughout the entire codebase. And second, it avoids disagreements about how a piece of code should be formatted.

Whenever you modify a file, run yapf on it to reformat it before checking it in.

yapf -i <modified file>

Yapf is run on every pull request to make sure the formatting is correct, so if you forget to do this the continuous integration system will remind you. Because different versions of yapf can produce different results, it is essential to use the same version that is being run on CI. At present, that is 0.22. We periodically update it to newer versions.


All classes and functions should include docstrings describing their purpose and intended usage. When in doubt about how much information to include, always err on the side of including more rather than less. Explain what problem a class is intended to solve, what algorithms it uses, and how to use it correctly. When appropriate, cite the relevant publications.

All docstrings should follow the numpy docstring formatting conventions.

Unit Tests

Having an extensive collection of test cases is essential to ensure the code works correctly. If you haven’t written tests for a feature, that means the feature isn’t finished yet. Untested code is code that probably doesn’t work.

Complex numerical code is sometimes challenging to fully test. When an algorithm produces a result, it sometimes is not obvious how to tell whether the result is correct or not. As far as possible, try to find simple examples for which the correct answer is exactly known. Sometimes we rely on stochastic tests which will probably pass if the code is correct and probably fail if the code is broken. This means these tests are expected to fail a small fraction of the time. Such tests can be marked with the @flaky annotation. If they fail during continuous integration, they will be run a second time and an error only reported if they fail again.

If possible, each test should run in no more than a few seconds. Occasionally this is not possible. In that case, mark the test with the @pytest.mark.slow annotation. Slow tests are skipped during continuous integration, so changes that break them may sometimes slip through and get merged into the repository. We still try to run them regularly, so hopefully the problem will be discovered fairly soon.

Type Annotations

Type annotations are an important tool for avoiding bugs. All new code should provide type annotations for function arguments and return types. When you make significant changes to existing code that does not have type annotations, please consider adding them at the same time.

We use the mypy static type checker to verify code correctness. It is automatically run on every pull request. If you want to run it locally to make sure you are using types correctly before checking in your code, cd to the top level directory of the repository and execute the command

mypy -p deepchem --ignore-missing-imports

Because Python is such a dynamic language, it sometimes is not obvious what type to specify. A good rule of thumb is to be permissive about input types and strict about output types. For example, many functions are documented as taking a list as an argument, but actually work just as well with a tuple. In those cases, it is best to specify the input type as Sequence to accept either one. But if a function returns a list, specify the type as List because we can guarantee the return value will always have that exact type.

Another important case is NumPy arrays. Many functions are documented as taking an array, but actually can accept any array-like object: a list of numbers, a list of lists of numbers, a list of arrays, etc. In that case, specify the type as Sequence to accept any of these. On the other hand, if the function truly requires an array and will fail with any other input, specify it as np.ndarray.

The deepchem.utils.typing module contains definitions of some types that appear frequently in the DeepChem API. You may find them useful when annotating code.