Description
System information
N/A
Describe the problem
Background
PEP 484 [1] added support for type hints in Python. These are purely annotations and are not enforced by the interpreter, however there are tools such as mypy [2] which can be run to check for consistency in the annotations. The typeshed initiative [3] has started to build external collections of type annotations for commonly used libraries.
When adding type annotations to a codebase, it is best if you can achieve near 100% coverage, otherwise uncertainty propagates out from everywhere the "untyped" code is called. A codebase using TF would likely struggle to gain much benefit from type-checking in any of the core code built on top of TF.
Benefits of Adding Type Annotations
- The expected inputs and outputs of functions become much clearer
- Code completion is able to provide more useful suggestions, boosting productivity by reducing amount of time spent referring to docs
- Static analysis can uncover latent bugs (case study here[5])
Difficulties/Drawbacks
- People may be encouraged to overly constrain types, removing some of the flexibility of a dynamic language. But given that Google's Python style-guide discourages "Power Features" [4] I would argue that striving towards code that is explicit is a similar philosophy
- The protobuf compiler would need to be augmented to generate type annotations.
- The Tensorflow Python codebase is huge, so at this point adding the annotations would be a huge undertaking.
- Tensorflow still supports python 2.7, 3.3 and 3.4 which do not have the type annotation syntax. So if this were implemented it would probably have to be in external *.pyi files, which is harder to maintain compared to inline type annotations in the source code.
Final thoughts
I realise that this would be a major undertaking and wouldn't be likely to ship any time soon, but I'm curious to gauge Google's thoughts on this new feature in Python. I'm about to start building a new codebase from scratch and was keen to use it as a chance to try out type annotations. I probably still will give it a shot, but I suspect that unless most of the common data science libs out there adopt this standard then its usefulness will be quite limited.
[1] https://www.python.org/dev/peps/pep-0484/
[2] http://mypy-lang.org/
[3] https://github.com/python/typeshed
[4] https://google.github.io/styleguide/pyguide.html#Power_Features
[5] http://blog.zulip.org/2016/10/13/static-types-in-python-oh-mypy/