Ariadne

This is the top level repository for Ariadne code. More information on using the Ariadne tools can be found here. This repository is code to analyze machine learning code with WALA. Currently, the code consists of the analysis of Python (com.ibm.wala.cast.python), analysis focused on machine learning in Python (com.ibm.wala.cast.python.ml), support for using the analysis via J2EE WebSockets (com.ibm.wala.cast.python.ml.j2ee) and their associated test projects.

Since it is built using WALA, you need to have WALA on your system to use it. Instructions on building this project can be found in CONTRIBUTING.md.

To test, for example, run TestCalls in the com.ibm.wala.cast.python.test project.

Type Annotation Sidecars

Tensor values loaded from data files (np.load, pickle.load, tf.io.read_file) have types the analysis cannot compute from code. A project may declare them in a sidecar file named ariadne-types.json at a Python-path root, or in the file named by the ariadne.typeAnnotations.file system property:

{
"types": [
	{
	"module": "datas/loader.py",
	"function": "Planetoid.load",
	"variable": "allx",
	"dtype": "float32",
	"shape": "1708 3703",
	"attribution": "Extent of the pickled Cora feature matrix."
	}
]
}

module is the root-relative script path; function is the dotted qualified name within the module, omitted or empty for module level; variable is the source-level name of the annotated binding. dtype and shape are each optional (an absent axis stays unknown). The shape is a whitespace-separated dimension string of integers or symbolic names; ... alone declares an unknown rank, and the empty string declares rank 0. The optional attribution cites the evidence that the value is content-dependent and is echoed in every report about the entry.

Annotations extend the analysis but never override it. An entry seeds a binding only where inference produced no type; a disagreement with an inferred type is reported, and the annotation is not applied. Entries that match no analyzed binding are also reported. Applied seeds carry a dedicated ANNOTATION provenance, keeping evidence that rests on user-supplied facts auditable downstream. See wala/ML#370.