metrics: plug the new operator function for fidelity

xplique/metrics/fidelity.py

@@ -9,7 +9,7 @@
 from scipy.stats import spearmanr
 
 from .base import ExplanationMetric
-from ..commons import batch_predictions_one_hot
+from ..commons import batch_predictions
 from ..types import Union, Callable, Optional, Dict
 
 
@@ -83,8 +83,8 @@ def __init__(self,
             (self.nb_samples, self.grid_size, self.grid_size, 1))
         self.subset_masks = tf.image.resize(subset_masks, inputs.shape[1:-1], method="nearest")
 
-        self.base_predictions = batch_predictions_one_hot(self.model, inputs,
-                                                          targets, self.batch_size)
+        self.base_predictions = batch_predictions(self.model, inputs,
+                                                  targets, self.batch_size)
 
     def evaluate(self,
                  explanations: Union[tf.Tensor, np.ndarray]) -> float:
@@ -117,8 +117,8 @@ def evaluate(self,
             # use the masks to set the selected subsets to baseline state
             degraded_inputs = inp * self.subset_masks + (1.0 - self.subset_masks) * baseline
             # measure the two terms that should be correlated
-            preds = base - batch_predictions_one_hot(self.model, degraded_inputs,
-                                                     label, self.batch_size)
+            preds = base - batch_predictions(self.model, degraded_inputs,
+                                            label, self.batch_size)
 
             attrs = tf.reduce_sum(phi * (1.0 - self.subset_masks), (1, 2, 3))
             corr_score = spearmanr(preds, attrs)[0]
@@ -276,8 +276,8 @@ def detailed_evaluate(self,
 
             batch_inputs = batch_inputs.reshape((-1, *self.inputs.shape[1:]))
 
-            predictions = batch_predictions_one_hot(self.model, batch_inputs,
-                                                    self.targets, self.batch_size)
+            predictions = batch_predictions(self.model, batch_inputs,
+                                            self.targets, self.batch_size)
 
             scores_dict[step] = np.mean(predictions)