Merge pull request #4985 from RasaHQ/bug-fix-response-selector

Fix errors during training in ResponseSelector

changelog/4985.bugfix.rst

@@ -0,0 +1 @@
+Fix errors during training and testing of ``ResponseSelector``.

rasa/nlu/classifiers/embedding_intent_classifier.py

@@ -79,7 +79,7 @@ class EmbeddingIntentClassifier(Component):
         "random_seed": None,
         # embedding parameters
         # default dense dimension used if no dense features are present
-        "dense_dim": {"text": 512, "intent": 20},
+        "dense_dim": {"text": 512, "label": 20},
         # dimension size of embedding vectors
         "embed_dim": 20,
         # the type of the similarity
@@ -137,12 +137,12 @@ def _check_old_config_variables(config: Dict[Text, Any]) -> None:
     def _load_nn_architecture_params(self, config: Dict[Text, Any]) -> None:
         self.hidden_layer_sizes = {
             "text": config["hidden_layers_sizes_a"],
-            "intent": config["hidden_layers_sizes_b"],
+            "label": config["hidden_layers_sizes_b"],
         }
         self.share_hidden_layers = config["share_hidden_layers"]
         if (
             self.share_hidden_layers
-            and self.hidden_layer_sizes["text"] != self.hidden_layer_sizes["intent"]
+            and self.hidden_layer_sizes["text"] != self.hidden_layer_sizes["label"]
         ):
             raise ValueError(
                 "If hidden layer weights are shared,"
@@ -307,15 +307,15 @@ def _extract_and_add_features(
         return sparse_features, dense_features
 
     def _extract_labels_precomputed_features(
-        self, label_examples: List["Message"]
+        self, label_examples: List["Message"], attribute: Text = INTENT_ATTRIBUTE
     ) -> List[np.ndarray]:
         """Collect precomputed encodings"""
 
         sparse_features = []
         dense_features = []
 
         for e in label_examples:
-            _sparse, _dense = self._extract_and_add_features(e, INTENT_ATTRIBUTE)
+            _sparse, _dense = self._extract_and_add_features(e, attribute)
             if _sparse is not None:
                 sparse_features.append(_sparse)
             if _dense is not None:
@@ -369,21 +369,23 @@ def _create_label_data(
 
         # Collect features, precomputed if they exist, else compute on the fly
         if self._check_labels_features_exist(labels_example, attribute):
-            features = self._extract_labels_precomputed_features(labels_example)
+            features = self._extract_labels_precomputed_features(
+                labels_example, attribute
+            )
         else:
             features = self._compute_default_label_features(labels_example)
 
         label_data = {}
-        self._add_to_session_data(label_data, "intent_features", features)
-        self._add_mask_to_session_data(label_data, "intent_mask", "intent_features")
+        self._add_to_session_data(label_data, "label_features", features)
+        self._add_mask_to_session_data(label_data, "label_mask", "label_features")
 
         return label_data
 
     def _use_default_label_features(self, label_ids: np.ndarray) -> List[np.ndarray]:
         return [
             np.array(
                 [
-                    self._label_data["intent_features"][0][label_id]
+                    self._label_data["label_features"][0][label_id]
                     for label_id in label_ids
                 ]
             )
@@ -394,7 +396,7 @@ def _create_session_data(
         self,
         training_data: List["Message"],
         label_id_dict: Optional[Dict[Text, int]] = None,
-        label_attribute: Optional[Text] = None,
+        label_attribute: Optional[Text] = INTENT_ATTRIBUTE,
     ) -> "SessionDataType":
         """Prepare data for training and create a SessionDataType object"""
 
@@ -405,20 +407,21 @@ def _create_session_data(
         label_ids = []
 
         for e in training_data:
-            _sparse, _dense = self._extract_and_add_features(e, TEXT_ATTRIBUTE)
-            if _sparse is not None:
-                X_sparse.append(_sparse)
-            if _dense is not None:
-                X_dense.append(_dense)
-
-            _sparse, _dense = self._extract_and_add_features(e, INTENT_ATTRIBUTE)
-            if _sparse is not None:
-                Y_sparse.append(_sparse)
-            if _dense is not None:
-                Y_dense.append(_dense)
-
-            if label_attribute and e.get(label_attribute):
-                label_ids.append(label_id_dict[e.get(label_attribute)])
+            if e.get(label_attribute):
+                _sparse, _dense = self._extract_and_add_features(e, TEXT_ATTRIBUTE)
+                if _sparse is not None:
+                    X_sparse.append(_sparse)
+                if _dense is not None:
+                    X_dense.append(_dense)
+
+                _sparse, _dense = self._extract_and_add_features(e, label_attribute)
+                if _sparse is not None:
+                    Y_sparse.append(_sparse)
+                if _dense is not None:
+                    Y_dense.append(_dense)
+
+                if label_id_dict:
+                    label_ids.append(label_id_dict[e.get(label_attribute)])
 
         X_sparse = np.array(X_sparse)
         X_dense = np.array(X_dense)
@@ -428,23 +431,21 @@ def _create_session_data(
 
         session_data = {}
         self._add_to_session_data(session_data, "text_features", [X_sparse, X_dense])
-        self._add_to_session_data(session_data, "intent_features", [Y_sparse, Y_dense])
+        self._add_to_session_data(session_data, "label_features", [Y_sparse, Y_dense])
         # explicitly add last dimension to label_ids
         # to track correctly dynamic sequences
         self._add_to_session_data(
-            session_data, "intent_ids", [np.expand_dims(label_ids, -1)]
+            session_data, "label_ids", [np.expand_dims(label_ids, -1)]
         )
 
-        if label_attribute and (
-            "intent_features" not in session_data or not session_data["intent_features"]
+        if label_id_dict and (
+            "label_features" not in session_data or not session_data["label_features"]
         ):
             # no label features are present, get default features from _label_data
-            session_data["intent_features"] = self._use_default_label_features(
-                label_ids
-            )
+            session_data["label_features"] = self._use_default_label_features(label_ids)
 
         self._add_mask_to_session_data(session_data, "text_mask", "text_features")
-        self._add_mask_to_session_data(session_data, "intent_mask", "intent_features")
+        self._add_mask_to_session_data(session_data, "label_mask", "label_features")
 
         return session_data
 
@@ -547,29 +548,29 @@ def _build_tf_train_graph(
             batch_data["text_features"], batch_data["text_mask"][0], "text"
         )
         b = self._combine_sparse_dense_features(
-            batch_data["intent_features"], batch_data["intent_mask"][0], "intent"
+            batch_data["label_features"], batch_data["label_mask"][0], "label"
         )
         all_bs = self._combine_sparse_dense_features(
-            label_data["intent_features"], label_data["intent_mask"][0], "intent"
+            label_data["label_features"], label_data["label_mask"][0], "label"
         )
 
         self.message_embed = self._create_tf_embed_fnn(
             a,
             self.hidden_layer_sizes["text"],
-            fnn_name="text_intent" if self.share_hidden_layers else "text",
+            fnn_name="text_label" if self.share_hidden_layers else "text",
             embed_name="text",
         )
         self.label_embed = self._create_tf_embed_fnn(
             b,
-            self.hidden_layer_sizes["intent"],
-            fnn_name="text_intent" if self.share_hidden_layers else "intent",
-            embed_name="intent",
+            self.hidden_layer_sizes["label"],
+            fnn_name="text_label" if self.share_hidden_layers else "label",
+            embed_name="label",
         )
         self.all_labels_embed = self._create_tf_embed_fnn(
             all_bs,
-            self.hidden_layer_sizes["intent"],
-            fnn_name="text_intent" if self.share_hidden_layers else "intent",
-            embed_name="intent",
+            self.hidden_layer_sizes["label"],
+            fnn_name="text_label" if self.share_hidden_layers else "label",
+            embed_name="label",
         )
 
         return train_utils.calculate_loss_acc(
@@ -606,15 +607,15 @@ def _build_tf_pred_graph(self, session_data: "SessionDataType") -> "tf.Tensor":
             batch_data["text_features"], batch_data["text_mask"][0], "text"
         )
         b = self._combine_sparse_dense_features(
-            batch_data["intent_features"], batch_data["intent_mask"][0], "intent"
+            batch_data["label_features"], batch_data["label_mask"][0], "label"
         )
 
         self.all_labels_embed = tf.constant(self.session.run(self.all_labels_embed))
 
         self.message_embed = self._create_tf_embed_fnn(
             a,
             self.hidden_layer_sizes["text"],
-            fnn_name="text_intent" if self.share_hidden_layers else "text",
+            fnn_name="text_label" if self.share_hidden_layers else "text",
             embed_name="text",
         )
 
@@ -626,9 +627,9 @@ def _build_tf_pred_graph(self, session_data: "SessionDataType") -> "tf.Tensor":
 
         self.label_embed = self._create_tf_embed_fnn(
             b,
-            self.hidden_layer_sizes["intent"],
-            fnn_name="text_intent" if self.share_hidden_layers else "intent",
-            embed_name="intent",
+            self.hidden_layer_sizes["label"],
+            fnn_name="text_label" if self.share_hidden_layers else "label",
+            embed_name="label",
         )
 
         self.sim = train_utils.tf_raw_sim(
@@ -651,7 +652,7 @@ def check_input_dimension_consistency(self, session_data: "SessionDataType"):
         if self.share_hidden_layers:
             num_text_features = self._get_num_of_features(session_data, "text_features")
             num_intent_features = self._get_num_of_features(
-                session_data, "intent_features"
+                session_data, "label_features"
             )
 
             if num_text_features != num_intent_features:
@@ -689,7 +690,7 @@ def preprocess_train_data(self, training_data: "TrainingData"):
 
     @staticmethod
     def _check_enough_labels(session_data: "SessionDataType") -> bool:
-        return len(np.unique(session_data["intent_ids"])) >= 2
+        return len(np.unique(session_data["label_ids"])) >= 2
 
     def train(
         self,
@@ -721,7 +722,7 @@ def train(
                 session_data,
                 self.evaluate_on_num_examples,
                 self.random_seed,
-                label_key="intent_ids",
+                label_key="label_ids",
             )
         else:
             eval_session_data = None
@@ -743,7 +744,7 @@ def train(
                 eval_session_data,
                 batch_size_in,
                 self.batch_in_strategy,
-                label_key="intent_ids",
+                label_key="label_ids",
             )
 
             self._is_training = tf.placeholder_with_default(False, shape=())

rasa/nlu/classifiers/sklearn_intent_classifier.py

@@ -5,7 +5,7 @@
 import typing
 from typing import Any, Dict, List, Optional, Text, Tuple
 
-from rasa.nlu.featurizers.featurzier import sequence_to_sentence_features
+from rasa.nlu.featurizers.featurizer import sequence_to_sentence_features
 from rasa.nlu import utils
 from rasa.nlu.classifiers import LABEL_RANKING_LENGTH
 from rasa.nlu.components import Component

rasa/nlu/featurizers/dense_featurizer/convert_featurizer.py

@@ -1,6 +1,6 @@
 import logging
 import warnings
-from rasa.nlu.featurizers.featurzier import Featurizer
+from rasa.nlu.featurizers.featurizer import Featurizer
 from typing import Any, Dict, List, Optional, Text, Tuple
 from rasa.nlu.config import RasaNLUModelConfig
 from rasa.nlu.training_data import Message, TrainingData

rasa/nlu/featurizers/dense_featurizer/mitie_featurizer.py

@@ -3,7 +3,7 @@
 from typing import Any, List, Text, Dict, Optional
 
 from rasa.nlu.config import RasaNLUModelConfig
-from rasa.nlu.featurizers.featurzier import Featurizer
+from rasa.nlu.featurizers.featurizer import Featurizer
 from rasa.nlu.tokenizers.tokenizer import Token
 from rasa.nlu.training_data import Message, TrainingData
 

rasa/nlu/featurizers/dense_featurizer/spacy_featurizer.py

@@ -3,7 +3,7 @@
 from typing import Any, Optional, Dict, Text
 
 from rasa.nlu.config import RasaNLUModelConfig
-from rasa.nlu.featurizers.featurzier import Featurizer
+from rasa.nlu.featurizers.featurizer import Featurizer
 from rasa.nlu.training_data import Message, TrainingData
 
 if typing.TYPE_CHECKING:

rasa/nlu/featurizers/sparse_featurizer/count_vectors_featurizer.py

@@ -7,7 +7,7 @@
 from sklearn.feature_extraction.text import CountVectorizer
 from rasa.nlu import utils
 from rasa.nlu.config import RasaNLUModelConfig
-from rasa.nlu.featurizers.featurzier import Featurizer
+from rasa.nlu.featurizers.featurizer import Featurizer
 from rasa.nlu.model import Metadata
 from rasa.nlu.training_data import Message, TrainingData
 from rasa.nlu.constants import (

rasa/nlu/featurizers/sparse_featurizer/ngram_featurizer.py

@@ -3,7 +3,7 @@
 
 from typing import Any, Dict, Optional, Text
 
-from rasa.nlu.featurizers.featurzier import Featurizer
+from rasa.nlu.featurizers.featurizer import Featurizer
 
 logger = logging.getLogger(__name__)
 

rasa/nlu/featurizers/sparse_featurizer/regex_featurizer.py

@@ -10,7 +10,7 @@
 
 from rasa.nlu import utils
 from rasa.nlu.config import RasaNLUModelConfig
-from rasa.nlu.featurizers.featurzier import Featurizer
+from rasa.nlu.featurizers.featurizer import Featurizer
 from rasa.nlu.training_data import Message, TrainingData
 import rasa.utils.io
 from rasa.nlu.constants import (

rasa/nlu/selectors/embedding_response_selector.py

@@ -43,7 +43,7 @@ class ResponseSelector(EmbeddingIntentClassifier):
 
     provides = ["response", "response_ranking"]
 
-    requires = [SPARSE_FEATURE_NAMES[TEXT_ATTRIBUTE]]
+    requires = []
 
     # default properties (DOC MARKER - don't remove)
     defaults = {
@@ -68,7 +68,7 @@ class ResponseSelector(EmbeddingIntentClassifier):
         "random_seed": None,
         # embedding parameters
         # default dense dimension used if no dense features are present
-        "dense_dim": 512,
+        "dense_dim": {"text": 512, "label": 20},
         # dimension size of embedding vectors
         "embed_dim": 20,
         # the type of the similarity

sample_configs/config_embedding_intent_response_selector.yml

@@ -0,0 +1,9 @@
+language: "en"
+
+pipeline:
+  - name: "WhitespaceTokenizer"
+  - name: "CountVectorsFeaturizer"
+  - name: "EmbeddingIntentClassifier"
+    epochs: 2
+  - name: "ResponseSelector"
+    epochs: 2

tests/nlu/featurizers/test_featurizer.py

@@ -2,7 +2,7 @@
 import pytest
 import scipy.sparse
 
-from rasa.nlu.featurizers.featurzier import Featurizer, sequence_to_sentence_features
+from rasa.nlu.featurizers.featurizer import Featurizer, sequence_to_sentence_features
 from rasa.nlu.constants import DENSE_FEATURE_NAMES, SPARSE_FEATURE_NAMES, TEXT_ATTRIBUTE
 from rasa.nlu.training_data import Message
 

tests/nlu/selectors/test_embedding_response_selector.py

@@ -0,0 +1,25 @@
+from rasa.nlu.training_data import load_data
+from rasa.nlu import config
+from rasa.nlu.train import Trainer, Interpreter
+
+
+def test_train_response_selector(component_builder, tmpdir):
+    td = load_data("data/examples/rasa/demo-rasa.md")
+    td_responses = load_data("data/examples/rasa/demo-rasa-responses.md")
+    td = td.merge(td_responses)
+    td.fill_response_phrases()
+
+    nlu_config = config.load(
+        "sample_configs/config_embedding_intent_response_selector.yml"
+    )
+
+    trainer = Trainer(nlu_config)
+    trainer.train(td)
+
+    persisted_path = trainer.persist(tmpdir)
+
+    assert trainer.pipeline
+    loaded = Interpreter.load(persisted_path, component_builder)
+    assert loaded.pipeline
+    assert loaded.parse("hello") is not None
+    assert loaded.parse("Hello today is Monday, again!") is not None