Are TFLite models supported for RISE Perturbation + DRISE

[ai2ys]

I would like to try "RISE Perturbation + DRISE" similar to the Notebook https://github.com/XAITK/xaitk-saliency/blob/master/examples/DRISE.ipynb from the examples, but with a TFLite model.

In the video on the website https://www.kitware.com/kitware-receives-honorable-mention-for-explainable-ai-toolkit/ is gets mentioned that TensorFlow models are supported by XAITK-saliency, but I can not find any information on this.

[Purg]

Hi! TF models are supported notionally in that a DetectImageObjects implementation could be written to wrap the use of something from TensorFlow (our adapter interface does not use any framework-specific primitives). We do not have any stock implementations that utilize TensorFlow that would be turn-key right now. Some of our notebooks examples (specifically this one) show the creation of an implementation of this interface. The referenced example still uses pytorch, but I hope the mechanism makes sense enough that you could adapt it for your TensorFlow use case.

You may have already deduced the following, but I'll mention it anyway just in case:

• The model modularity in the DRISE example notebook happens in the blackbox_detector slot in the sal_map_generator.generate(ref_image, ref_bboxes, ref_scores, blackbox_detector) call.
• We define object-detection-oriented interface (here) that we expect the input to be an implementation of the DetectImageObjects, which is an adapter pattern interface defined in the smqtk_detection package (specific interface here).
• If you create a new implementation for DetectImageObjects that you think might be generally useful, the smqtk-detection package will happily accept contributions! (I/we also maintain that.)