Solve issues #150 and #151 on fidelity metrics

setup.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 1.3.2
+current_version = 1.3.3
 commit = True
 tag = False
 

setup.py

@@ -5,7 +5,7 @@
 
 setup(
     name="Xplique",
-    version="1.3.2",
+    version="1.3.3",
     description="Explanations toolbox for Tensorflow 2",
     long_description=README,
     long_description_content_type="text/markdown",

tests/metrics/test_common.py

@@ -112,7 +112,7 @@ def test_data_types_shapes():
     explanation_metrics = {
         Deletion: {"steps": 3},
         Insertion: {"steps": 3},
-        MuFidelity: {"nb_samples": 3},
+        MuFidelity: {"nb_samples": 3, "grid_size": None, "subset_percent": 0.9},
     }
 
     for data_type, input_shape in data_types_input_shapes.items():

tests/metrics/test_fidelity.py

@@ -28,12 +28,12 @@ def test_mu_fidelity():
 
 def test_causal_metrics():
     # ensure we can compute insertion/deletion metric with consistent arguments
-    input_shape, nb_labels, nb_samples = ((32, 32, 3), 10, 20)
+    input_shape, nb_labels, nb_samples = ((10, 10, 3), 10, 20)
     x, y = generate_data(input_shape, nb_labels, nb_samples)
     model = generate_model(input_shape, nb_labels)
     explanations = np.random.uniform(0, 1, x.shape[:-1] + (1,)).astype(np.float32)
 
-    for step in [5, 10]:
+    for step in [-1, 5, 10]:
         for baseline_mode in [0.0, lambda x: x-0.5]:
             score_insertion = Insertion(model, x, y,
                                         baseline_mode=baseline_mode,

xplique/__init__.py

@@ -6,7 +6,7 @@
 techniques
 """
 
-__version__ = '1.3.2'
+__version__ = '1.3.3'
 
 from . import attributions
 from . import concepts

xplique/metrics/fidelity.py

@@ -87,8 +87,6 @@ def __init__(self,
 
         # if unspecified use the original equation (pixel-wise modification)
         self.grid_size = grid_size or self.inputs.shape[1]
-        # cardinal of subset (|S| in the equation)
-        self.subset_size = int(self.grid_size ** 2 * self.subset_percent)
 
         self.base_predictions = self.batch_inference_function(self.model, self.inputs,
                                                               self.targets, self.batch_size)
@@ -198,14 +196,14 @@ def _perturb_samples(self,
         if len(inputs.shape) == 2:  # tabular data, grid size is ignored
             # prepare the random masks
             subset_masks = tf.random.uniform((nb_perturbations, inputs.shape[1]), 0, 1, tf.float32)
-            subset_masks = tf.argsort(subset_masks, axis=-1) > self.subset_size
+            subset_masks = subset_masks > self.subset_percent
             subset_masks = tf.cast(subset_masks, tf.float32)
 
         elif len(inputs.shape) == 3:  # time series
             # prepare the random masks
             subset_masks = tf.random.uniform((nb_perturbations, self.grid_size * inputs.shape[2]),
                                              minval=0, maxval=1, dtype=tf.float32)
-            subset_masks = tf.argsort(subset_masks, axis=-1) > self.subset_size
+            subset_masks = subset_masks > self.subset_percent
 
             # and interpolate them if needed
             subset_masks = tf.reshape(tf.cast(subset_masks, tf.float32),
@@ -217,7 +215,7 @@ def _perturb_samples(self,
             # prepare the random masks
             subset_masks = tf.random.uniform(shape=(nb_perturbations, self.grid_size ** 2),
                                              minval=0, maxval=1, dtype=tf.float32)
-            subset_masks = tf.argsort(subset_masks, axis=-1) > self.subset_size
+            subset_masks = subset_masks > self.subset_percent
 
             # and interpolate them if needed
             subset_masks = tf.reshape(tf.cast(subset_masks, tf.float32),
@@ -303,7 +301,7 @@ def __init__(self,
 
         assert 0.0 < max_percentage_perturbed <= 1.0, \
             "`max_percentage_perturbed` must be in ]0, 1]."
-        self.max_nb_perturbed = tf.math.floor(self.nb_features * max_percentage_perturbed)
+        self.max_nb_perturbed = int(np.floor(self.nb_features * max_percentage_perturbed))
 
         if steps == -1:
             steps = self.max_nb_perturbed