set values in MMImageMaskInput (#1675)

Summary:
Pull Request resolved: #1675

previously `InterpretableInput` is only used for text, the `values` are the actual text segments, which are used in `AttributionResult` for plot legends.

This design no longer makes sense for image. Will redesign the `values`. But for now, just set `values` for plot legends

Reviewed By: craymichael

Differential Revision: D87891825

fbshipit-source-id: 5b8e739f44d7b5983cdeff609a38cb53b48c7c88

Author: aobo-y

captum/attr/_utils/interpretable_input.py

@@ -572,7 +572,7 @@ class MMImageMaskInput(InterpretableInput):
     n_itp_features: int
     original_model_inputs: Any
     mask_id_to_idx: Dict[int, int]
-    values: List[str] = []  # no use for now
+    values: List[str]
 
     def __init__(
         self,
@@ -606,6 +606,10 @@ def __init__(
 
         self.original_model_inputs = processor_fn(image)
 
+        # temporarily for compatibility with AttributionResult
+        # which use the values for plot legends
+        self.values = [f"image_feature_{mid}" for mid in mask_ids]
+
     def to_tensor(self) -> Tensor:
         return torch.tensor([[1.0] * self.n_itp_features])