Merge pull request RasaHQ#6054 from RasaHQ/share-hidden-layers

Fix option `share_hidden_layers`

Author: rasabot

changelog/6053.bugfix.rst

@@ -0,0 +1,2 @@
+Update ``FeatureSignature`` to store just the feature dimension instead of the complete shape. This change fixes the
+usage of the option ``share_hidden_layers`` in the ``DIETClassifier``.

rasa/nlu/classifiers/diet_classifier.py

@@ -1183,14 +1183,14 @@ def _prepare_sparse_dense_layers(
     ) -> None:
         sparse = False
         dense = False
-        for is_sparse, shape in feature_signatures:
+        for is_sparse, feature_dimension in feature_signatures:
             if is_sparse:
                 sparse = True
             else:
                 dense = True
                 # if dense features are present
                 # use the feature dimension of the dense features
-                dense_dim = shape[-1]
+                dense_dim = feature_dimension
 
         if sparse:
             self._tf_layers[f"sparse_to_dense.{name}"] = layers.DenseForSparse(

rasa/utils/tensorflow/model_data.py

@@ -25,7 +25,7 @@ class FeatureSignature(NamedTuple):
     """Stores the shape and the type (sparse vs dense) of features."""
 
     is_sparse: bool
-    shape: List[int]
+    feature_dimension: Optional[int]
 
 
 class RasaModelData:
@@ -210,7 +210,7 @@ def get_signature(self) -> Dict[Text, List[FeatureSignature]]:
             key: [
                 FeatureSignature(
                     True if isinstance(v[0], scipy.sparse.spmatrix) else False,
-                    v[0].shape,
+                    v[0].shape[-1] if v[0].shape else None,
                 )
                 for v in values
             ]
@@ -357,8 +357,8 @@ def _balanced_data(self, data: Data, batch_size: int, shuffle: bool) -> Data:
                 if num_data_cycles[index] > 0 and not skipped[index]:
                     skipped[index] = True
                     continue
-                else:
-                    skipped[index] = False
+
+                skipped[index] = False
 
                 index_batch_size = (
                     int(counts_label_ids[index] / self.num_examples * batch_size) + 1

rasa/utils/tensorflow/models.py

@@ -417,15 +417,15 @@ def batch_to_model_data_format(
 
         idx = 0
         for k, signature in data_signature.items():
-            for is_sparse, shape in signature:
+            for is_sparse, feature_dimension in signature:
                 if is_sparse:
                     # explicitly substitute last dimension in shape with known
                     # static value
                     batch_data[k].append(
                         tf.SparseTensor(
                             batch[idx],
                             batch[idx + 1],
-                            [batch[idx + 2][0], batch[idx + 2][1], shape[-1]],
+                            [batch[idx + 2][0], batch[idx + 2][1], feature_dimension],
                         )
                     )
                     idx += 3