data unrolling: better handling of batched tf.Dataset

tests/attributions/test_common.py

@@ -57,6 +57,9 @@ def test_common():
             # all explanations returned must be either a tf.Tensor or ndarray
             assert isinstance(explanations, (tf.Tensor, np.ndarray))
 
+            # we should have one explanation for each inputs
+            assert len(explanations) == len(inputs_np)
+
 
 def test_batch_size():
     """Ensure the functioning of attributions for special batch size cases"""
@@ -89,12 +92,13 @@ def test_batch_size():
         ]
 
         for method in methods:
-            try:
-                explanations = method.explain(inputs, targets)
-            except:
-                raise AssertionError(
-                    "Explanation failed for method ", method.__class__.__name__,
-                    " batch size ", bs)
+            explanations = method.explain(inputs, targets)
+
+            # all explanations returned must be either a tf.Tensor or ndarray
+            assert isinstance(explanations, (tf.Tensor, np.ndarray))
+
+            # we should have one explanation for each inputs
+            assert len(explanations) == len(inputs)
 
 
 def test_model_caching():
@@ -118,4 +122,4 @@ def test_model_caching():
 
     # ensure that there no more than one key has been added
     assert (len(
-        BlackBoxExplainer._cache_models) == cache_len_before + 1)  # pylint: disable=protected-access
+        BlackBoxExplainer._cache_models) == cache_len_before + 1)  # pylint: disable=protected-access

xplique/commons/data_conversion.py

@@ -31,15 +31,13 @@ def tensor_sanitize(inputs: Union[tf.data.Dataset, tf.Tensor, np.ndarray],
 
     # deal with tf.data.Dataset
     if isinstance(inputs, tf.data.Dataset):
-        # if the dataset as 4 dimensions, assume it is batched
-        dataset_shape = inputs.element_spec[0].shape
-        if len(dataset_shape) == 4:
+        # try to know if the dataset is batched, if it is the case we unbatch
+        if hasattr(inputs, '_batch_size'):
             inputs = inputs.unbatch()
         # unpack the dataset, assume we have tuple of (input, target)
-        targets = [target for inp, target in inputs]
-        inputs  = [inp for inp, target in inputs]
+        targets = [target for _, target in inputs]
+        inputs  = [inp for inp, _ in inputs]
 
-    # deal with numpy array
     inputs = tf.cast(inputs, tf.float32)
     targets = tf.cast(targets, tf.float32)