Runtime type-driven synthesis based on domain specific language (DSL) transformations.

Development: Orestis Farmakis (https://github.com/TheSolipsist), Manios Krasanakis ([email protected])
Dependencies: scipy,torch,networkx,TODO

🚀 Run code only once
🤖 DSL rules from type-hinted methods
✍️ Custom type constraints

To use the library, lets start from a couple of methods will full type hints:

import numpy as np

def list2array(x: list) -> np.ndarray:
    return np.array(x)

def scale(x: np.ndarray, factor: "positive") -> np.ndarray:
    return x*factor

The "positive" hint is not declared yet, but that's alright. We now convert these methods to a pymeleon DSL, assign a name to it, and pass it to a GeneticViewer that will support type-driven synthesis.

from pymeleon import DSL, GeneticViewer, Predicate, autorule

viewer = DSL(
    Predicate("positive", lambda x: isinstance(x, float) and x > 0),
    autorule(list2array),
    autorule(scale)
).set_name("demo_dsl") >> GeneticViewer()

The DSL definition comprises any number of custom predicates (i.e., that are not python types and are passed as strings to typehints) and any number of DSL rules, which are transformations with specific input and output types. Type combinations are supported via custom rule generation .

Finally, if you can provide your DSL viewer to users of your code so that they can transform collections of objects to a specific type. They can do so by applying transformations on the objects per:

scale = viewer([1, 2, 3], 3.) >> np.ndarray
print(scale)
# [3. 6. 9.]

Again, more complex output requirements can be declared.

DSL Definition
Data Viewers