invert dialogue time in ted

changelog/7867.improvement.md

@@ -0,0 +1,3 @@
+If `MaxHistoryTrackerFeaturizer` is used, invert the dialogue sequence before passing
+it to the transformer so that the last dialogue input becomes the first one and
+therefore always have the same pos encoding.

rasa/core/policies/ted_policy.py

@@ -191,7 +191,7 @@ class TEDPolicy(Policy):
         # Max position for relative embeddings
         MAX_RELATIVE_POSITION: None,
         # Use a unidirectional or bidirectional encoder.
-        UNIDIRECTIONAL_ENCODER: True,
+        UNIDIRECTIONAL_ENCODER: False,
         # ## Training parameters
         # Initial and final batch sizes:
         # Batch size will be linearly increased for each epoch.
@@ -803,9 +803,9 @@ def load(
             model_data_example,
             data_signature=model_data_example.get_signature(),
             config=meta,
-            # during prediction we don't care about previous dialogue turns,
-            # so to save computation time, use only the last one
-            use_only_last_dialogue_turns=True,
+            use_only_last_dialogue_turns=isinstance(
+                featurizer, MaxHistoryTrackerFeaturizer
+            ),
             label_data=label_data,
             entity_tag_specs=entity_tag_specs,
             finetune_mode=should_finetune,
@@ -925,6 +925,11 @@ def _prepare_layers(self) -> None:
             self.config[TRANSFORMER_SIZE][DIALOGUE],
             self.config[DROP_RATE_DIALOGUE],
             self.config[DROP_RATE_ATTENTION],
+            # use bidirectional transformer, because
+            # we will invert dialogue sequence so that the last turn is located
+            # at the first position and would always have
+            # exactly the same positional encoding
+            unidirectional=not self.use_only_last_dialogue_turns,
         )
 
         self._prepare_embed_layers(DIALOGUE)
@@ -1062,13 +1067,24 @@ def _embed_dialogue(
         dialogue_lengths = tf.cast(tf_batch_data[DIALOGUE][LENGTH][0], tf.int32)
         mask = self._compute_mask(dialogue_lengths)
 
+        if self.use_only_last_dialogue_turns:
+            # invert dialogue sequence so that the last turn would always have
+            # exactly the same positional encoding
+            dialogue_in = tf.reverse_sequence(dialogue_in, dialogue_lengths, seq_axis=1)
+
         dialogue_transformed, attention_weights = self._tf_layers[
             f"transformer.{DIALOGUE}"
         ](dialogue_in, 1 - mask, self._training)
         dialogue_transformed = tfa.activations.gelu(dialogue_transformed)
 
         if self.use_only_last_dialogue_turns:
-            # pick last vector if max history featurizer is used
+            # pick last vector if max history featurizer is used, since we inverted
+            # dialogue sequence, the last vector is actually the first one
+            dialogue_transformed = dialogue_transformed[:, :1, :]
+            mask = tf.expand_dims(self._last_token(mask, dialogue_lengths), 1)
+        elif not self._training:
+            # during prediction we don't care about previous dialogue turns,
+            # so to save computation time, use only the last one
             dialogue_transformed = tf.expand_dims(
                 self._last_token(dialogue_transformed, dialogue_lengths), 1
             )
@@ -1234,7 +1250,7 @@ def _encode_real_features_per_attribute(
         Args:
             tf_batch_data: dictionary mapping every attribute to its features and masks
             attribute: the attribute we will encode features for
-            (e.g., ACTION_NAME, INTENT)
+                (e.g., ACTION_NAME, INTENT)
 
         Returns:
             A tensor combining  all features for `attribute`
@@ -1334,10 +1350,9 @@ def _convert_to_original_shape(
 
         Args:
             attribute_features: the "real" features to convert
-            attribute_mask:  the tensor containing the position of "real" features
-                in the dialogue, shape is (batch-size x dialogue_len x 1)
-            dialogue_lengths: the tensor containing the actual dialogue length,
-                shape is (batch-size,)
+            tf_batch_data: dictionary mapping every attribute to its features and masks
+            attribute: the attribute we will encode features for
+                (e.g., ACTION_NAME, INTENT)
 
         Returns:
             The converted attribute features

rasa/utils/tensorflow/models.py

@@ -755,6 +755,7 @@ def _prepare_transformer_layer(
         units: int,
         drop_rate: float,
         drop_rate_attention: float,
+        unidirectional: bool,
         prefix: Text = "transformer",
     ):
         if num_layers > 0:
@@ -767,7 +768,7 @@ def _prepare_transformer_layer(
                 dropout_rate=drop_rate,
                 attention_dropout_rate=drop_rate_attention,
                 sparsity=self.config[WEIGHT_SPARSITY],
-                unidirectional=self.config[UNIDIRECTIONAL_ENCODER],
+                unidirectional=unidirectional,
                 use_key_relative_position=self.config[KEY_RELATIVE_ATTENTION],
                 use_value_relative_position=self.config[VALUE_RELATIVE_ATTENTION],
                 max_relative_position=self.config[MAX_RELATIVE_POSITION],
@@ -872,6 +873,7 @@ def _prepare_sequence_layers(self, name: Text) -> None:
             size,
             self.config[DROP_RATE],
             self.config[DROP_RATE_ATTENTION],
+            self.config[UNIDIRECTIONAL_ENCODER],
         )
 
     def _prepare_entity_recognition_layers(self) -> None: