jacereda/python-deduckt

Runtime type inference for Python

python-deduckt

Through the powers of deductive interpretation, Deduckt will obtain the precise call graph of your Python programs and the type signatures of all executed functions. Created for the goals of py2nim

How to use

python-deduckt (also launched with python-deduckt/deduckt/main.py) is a drop-in replacement for the python interpreter.

Each traced python execution will update the python-deduckt.json file stored in the current directory. This file stores an annotated AST for each module imported by your program.

python-deduckt full path test.py args

Results

The format is

{
    "<pythonName>": {
        <typeAnnotation>
    }
}

Python name

"<namespaces>.<name>"

name

"<name>" # const
"<name>" # class
"#<name>" # function
"<class>#<name>" # method

Variables

{
    "name": <name>,
    "type": <typeAnnotation>
}

A type annotation can be a class, function description or an atomic type.

{
    "kind": "PyTypeFunction",
    "label": <name>,
    "args": [<variable>],
    "variables": [<variable>],
    "returnType": <typeAnnotation>
}

{
    "kind": "PyTypeObject",
    "label": <name>,
    "fields": [<variable>]
}

Full doc: todo

FAQ

• Can I generate mypy annotations with it?

It's definitely possible to generate very useful mypy annotations from the recorded data. However that hasn't been our use case, but contributions are welcome

• Why does it analyze all the events?

In the future we might add an option to trace only a statistically significant part of them. It's not a huge priority for our use case, as we used it for py2nim. A porting task is not something that would be ran often and better type information is more important than speed.

• How to pronounce deduckt?

Like the first part of "deductive", or like getting rid of the duck types in your backyard.

LICENSE

The MIT License (MIT)

Copyright (c) 2017-2018 Zahary Karadjov, Alexander Ivanov