[GSoC] 6. Enabling full OpenVINO support in Anomalib

[samet-akcay]

Short description

Anomalib currently supports fundamental OpenVINO features, including the export of trained models to OpenVINO Intermediate Representation (IR). This project would expand the OpenVINO export capabilities of Anomalib. In general, the milestones would have the following structure:

• Create an export module with multiple CLI arguments that configure the type of optimization. (e.g., POT, MO, and NNCF, as well as advanced settings for each)
• Enable sync and async inference modes for OpenVINO in anomalib.
• Integrate OpenVINO's new Python API with Anomalib's OpenVINO interface, which currently utilizes the inference engine, to be deprecated in future releases.

NOTE: a portion of the workload may overlap with another project, as both require implementing the async inference capabilities of OpenVINO. This would be a great opportunity to learn how to collaborate with other developers and resolve git merge conflicts, which are common in everyday software development.

Expected outcomes

An export module in anomalib that converts anomalib models into OpenVINO via POT, MO, or NNCF, OpenVINO async inference interface using OpenVINO’s new Python API.

Skills required/preferred

Python, OpenVINO (intermediate)

Mentors

• Samet Akcay
• Ashwin Vaidya

Size of project

• 350 hours

Difficulty

• Medium