Change defaults in EmbeddingIntentClassifier and DIETClassifier

changelog/5717.improvement.rst

@@ -0,0 +1,3 @@
+Change default values for hyper-parameters in ``EmbeddingIntentClassifier`` and ``DIETClassifier``
+
+Use ``scale_loss=False`` in ``DIETClassifier``. Reduce the number of dense dimensions for sparse features of text from 512 to 256 in ``EmbeddingIntentClassifier``.

docs/nlu/components.rst

@@ -883,7 +883,7 @@ EmbeddingIntentClassifier
          +---------------------------------+------------------+--------------------------------------------------------------+
          | 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: 256        | Dense dimension for sparse features to use if no dense       |
          |                                 | label: 20        | features are present.                                        |
          +---------------------------------+------------------+--------------------------------------------------------------+
          | number_of_negative_examples     | 20               | The number of incorrect labels. The algorithm will minimize  |
@@ -1702,7 +1702,7 @@ DIETClassifier
          |                                 |                  | over incorrect intent labels, used only if 'loss_type' is    |
          |                                 |                  | set to 'margin'.                                             |
          +---------------------------------+------------------+--------------------------------------------------------------+
-         | scale_loss                      | True             | Scale loss inverse proportionally to confidence of correct   |
+         | scale_loss                      | False            | Scale loss inverse proportionally to confidence of correct   |
          |                                 |                  | prediction.                                                  |
          +---------------------------------+------------------+--------------------------------------------------------------+
          | regularization_constant         | 0.002            | The scale of regularization.                                 |

pyproject.toml

@@ -9,7 +9,7 @@ exclude = "((.eggs | .git | .pytype | .pytest_cache | build | dist))"
 
 [tool.poetry]
 name = "rasa"
-version = "1.9.7"
+version = "1.10.0a3"
 description = "Open source machine learning framework to automate text- and voice-based conversations: NLU, dialogue management, connect to Slack, Facebook, and more - Create chatbots and voice assistants"
 authors = [ "Rasa Technologies GmbH <hi@rasa.com>",]
 maintainers = [ "Tom Bocklisch <tom@rasa.com>",]

rasa/nlu/classifiers/diet_classifier.py

@@ -189,7 +189,7 @@ def required_components(cls) -> List[Type[Component]]:
         USE_MAX_NEG_SIM: True,
         # If 'True' scale loss inverse proportionally to the confidence
         # of the correct prediction
-        SCALE_LOSS: True,
+        SCALE_LOSS: False,
         # ## Regularization parameters
         # The scale of regularization
         REGULARIZATION_CONSTANT: 0.002,

rasa/nlu/classifiers/embedding_intent_classifier.py

@@ -93,7 +93,7 @@ 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},
         # The number of incorrect labels. The algorithm will minimize
         # their similarity to the user input during training.
         NUM_NEG: 20,

rasa/version.py

@@ -1,3 +1,3 @@
 # this file will automatically be changed,
 # do not add anything but the version number here!
-__version__ = "1.9.7"
+__version__ = "1.10.0a3"