Insertion Deletion AUC metric

[GalyaZalesskaya]

• Added Insertion Deletion AUC metric that calculates the accuracy drop/increase during deletion/insertion of important pixels
  

• Added parent BasicMetric

[codecov]

Codecov Report

Attention: Patch coverage is 98.73418% with 1 line in your changes missing coverage. Please review.

Project coverage is 92.64%. Comparing base (1e41ff2) to head (dab6e9a).
Report is 1 commits behind head on develop.

Files	Patch %	Lines
openvino_xai/explainer/utils.py	88.88%	1 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop      #56      +/-   ##
===========================================
+ Coverage    92.43%   92.64%   +0.21%     
===========================================
  Files           20       22       +2     
  Lines         1308     1373      +65     
===========================================
+ Hits          1209     1272      +63     
- Misses          99      101       +2     

Flag	Coverage Δ
dev-py310	92.64% <98.73%> (+0.21%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[goodsong81]

Thank you for the update. I have a few comments below.
Besides, one alternative design idea came up in my mind. Could you please consider the following idea and my comments?

class Metric(ABC):
    def __init__(
        self,
        score_fn: Callable[[np.ndarray], np.ndarray],  # image -> score vector
        # metric specific static parameters (like 'step' in AUC)
        *args,
        **kwargs,
    ):
        ...

    def __call__(  # Single image
        self,
        saliency_map: np.ndarray,
        class_idx: int,
        input_image: np.ndarray,
    ) -> Dict[str, float]:
        ...

    def evaluate(  # Dataset
        self,
        explanations: List[Explanation],
        input_images: List[np.ndarray],
    ) -> Dict[str, float]:
        ...

To me, setting model, preproc, postproc functions looks too complicated.
In above way, users could set arbitrary inference function that computes score vector from image. Explainer.model_forward() could also be reused in this way.

[GalyaZalesskaya]

Thank you for your suggestions, Songki. I implemented __call__ for a single image and evaluate for a set of images. I also return Dict[str, float] for both of functions. It helps the generality but getting the score is no more elegant:

pointing_game_score = self.pointing_game.evaluate([explanation], self.gt_bboxes)["pointing_game"]
pointing_game_score = list(self.pointing_game.evaluate([explanation], self.gt_bboxes).values())[0]

Provably, evaluate can affect the internal state of Metric class and we can get scores as property function pointing_game.score. Let's leave the optimization until the next PR.

[goodsong81]

Let's do the refactoring afterwards. Thank you for the update!