Merge 8ca8344 into ba0d890

RasaHQ · Sep 4, 2020 · 73f5d12 · 73f5d12
2 parents ba0d890 + 8ca8344
commit 73f5d12
Show file tree

Hide file tree

Showing 5 changed files with 14 additions and 13 deletions.
diff --git a/changelog/6555.improvement.md b/changelog/6555.improvement.md
@@ -0,0 +1,3 @@
+Do not set the output dimension of the `sparse-to-dense` layers to the same dimension as the dense features.
+
+Update default value of `dense_dimension` for `DIETClassifier` and `ResponseSelctor` to 128.
diff --git a/docs/docs/components/intent-classifiers.mdx b/docs/docs/components/intent-classifiers.mdx
@@ -364,7 +364,7 @@ Intent classifiers assign one of the intents defined in the domain file to incom
     +---------------------------------+------------------+--------------------------------------------------------------+
     | embedding_dimension             | 20               | Dimension size of embedding vectors.                         |
     +---------------------------------+------------------+--------------------------------------------------------------+
-    | dense_dimension                 | text: 512        | Dense dimension for sparse features to use if no dense       |
+    | dense_dimension                 | text: 128        | Dense dimension for sparse features to use if no dense       |
     |                                 | label: 20        | features are present.                                        |
     +---------------------------------+------------------+--------------------------------------------------------------+
     | concat_dimension                | text: 512        | Concat dimension for sequence and sentence features.         |

diff --git a/docs/docs/components/selectors.mdx b/docs/docs/components/selectors.mdx
@@ -187,9 +187,9 @@ Selectors predict a bot response from a set of candidate responses.
     +---------------------------------+-------------------+--------------------------------------------------------------+
     | embedding_dimension             | 20                | Dimension size of embedding vectors.                         |
     +---------------------------------+-------------------+--------------------------------------------------------------+
-    | dense_dimension                 | text: 512         | Dense dimension for sparse features to use if no dense       |
-    |                                 | label: 512        | features are present.                                        |
-    +---------------------------------+-------------------+--------------------------------------------------------------+
+    | dense_dimension                 | text: 128         | Dense dimension for sparse features to use if no dense       |
+    |                                 | label: 128        | features are present.                                        |
+    +---------------------------------+------------------+--------------------------------------------------------------+
     | concat_dimension                | text: 512         | Concat dimension for sequence and sentence features.         |
     |                                 | label: 512        |                                                              |
     +---------------------------------+-------------------+--------------------------------------------------------------+

diff --git a/rasa/nlu/classifiers/diet_classifier.py b/rasa/nlu/classifiers/diet_classifier.py
@@ -53,7 +53,6 @@
     EPOCHS,
     RANDOM_SEED,
     LEARNING_RATE,
-    DENSE_DIMENSION,
     RANKING_LENGTH,
     LOSS_TYPE,
     SIMILARITY_TYPE,
@@ -87,6 +86,7 @@
     TENSORBOARD_LOG_LEVEL,
     CONCAT_DIMENSION,
     FEATURIZERS,
+    DENSE_DIMENSION,
 )
 
 
@@ -174,9 +174,9 @@ def required_components(cls) -> List[Type[Component]]:
         # Dimension size of embedding vectors
         EMBEDDING_DIMENSION: 20,
         # Default dense dimension to use if no dense features are present.
-        DENSE_DIMENSION: {TEXT: 512, LABEL: 20},
+        DENSE_DIMENSION: {TEXT: 256, LABEL: 20},
         # Default dimension to use for concatenating sequence and sentence features.
-        CONCAT_DIMENSION: {TEXT: 512, LABEL: 20},
+        CONCAT_DIMENSION: {TEXT: 256, LABEL: 20},
         # The number of incorrect labels. The algorithm will minimize
         # their similarity to the user input during training.
         NUM_NEG: 20,
@@ -1340,14 +1340,12 @@ def _prepare_sparse_dense_layers(
     ) -> None:
         sparse = False
         dense = False
+
         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 = feature_dimension
 
         if sparse:
             self._tf_layers[f"sparse_to_dense.{name}"] = layers.DenseForSparse(

diff --git a/rasa/nlu/selectors/response_selector.py b/rasa/nlu/selectors/response_selector.py
@@ -36,7 +36,6 @@
     EPOCHS,
     RANDOM_SEED,
     LEARNING_RATE,
-    DENSE_DIMENSION,
     RANKING_LENGTH,
     LOSS_TYPE,
     SIMILARITY_TYPE,
@@ -72,6 +71,7 @@
     TENSORBOARD_LOG_LEVEL,
     CONCAT_DIMENSION,
     FEATURIZERS,
+    DENSE_DIMENSION,
 )
 from rasa.nlu.constants import (
     RESPONSE,
@@ -157,9 +157,9 @@ def required_components(cls) -> List[Type[Component]]:
         # Dimension size of embedding vectors
         EMBEDDING_DIMENSION: 20,
         # Default dense dimension to use if no dense features are present.
-        DENSE_DIMENSION: {TEXT: 512, LABEL: 512},
+        DENSE_DIMENSION: {TEXT: 256, LABEL: 20},
         # Default dimension to use for concatenating sequence and sentence features.
-        CONCAT_DIMENSION: {TEXT: 512, LABEL: 512},
+        CONCAT_DIMENSION: {TEXT: 256, LABEL: 256},
         # The number of incorrect labels. The algorithm will minimize
         # their similarity to the user input during training.
         NUM_NEG: 20,
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,3 @@
		Do not set the output dimension of the `sparse-to-dense` layers to the same dimension as the dense features.

		Update default value of `dense_dimension` for `DIETClassifier` and `ResponseSelctor` to 128.