def _prepare_layers(self) -> None:
self.text_name = TEXT
self.label_name = TEXT if self.config[SHARE_HIDDEN_LAYERS] else LABEL
# For user text and response text, prepare layers that combine different feature
# types, embed everything using a transformer and optionally also do masked
# language modeling. Omit input dropout for label features.
label_config = self.config.copy()
label_config.update({
SPARSE_INPUT_DROPOUT: False,
DENSE_INPUT_DROPOUT: False
})
for attribute, config in [
(self.text_name, self.config),
(self.label_name, label_config),
]:
self._tf_layers[
f"sequence_layer.{attribute}"] = rasa_layers.RasaSequenceLayer(
attribute, self.data_signature[attribute], config)
if self.config[MASKED_LM]:
self._prepare_mask_lm_loss(self.text_name)
self._prepare_label_classification_layers(
predictor_attribute=self.text_name)
def _prepare_layers(self) -> None:
self.text_name = TEXT
self.label_name = TEXT if self.config[SHARE_HIDDEN_LAYERS] else LABEL
# For user text and response text, prepare layers that combine different feature
# types, embed everything using a transformer and optionally also do masked
# language modeling.
for attribute in [self.text_name, self.label_name]:
self._tf_layers[
f"sequence_layer.{attribute}"] = rasa_layers.RasaSequenceLayer(
attribute, self.data_signature[attribute], self.config)
if self.config[MASKED_LM]:
self._prepare_mask_lm_loss(self.text_name)
self._prepare_label_classification_layers(
predictor_attribute=self.text_name)