async def test_process_gives_diagnostic_data(
create_train_load_and_process_diet: Callable[..., Message],
default_execution_context: ExecutionContext,
should_add_diagnostic_data: bool,
):
default_execution_context.should_add_diagnostic_data = should_add_diagnostic_data
default_execution_context.node_name = "DIETClassifier_node_name"
processed_message = create_train_load_and_process_diet({EPOCHS: 1})
if should_add_diagnostic_data:
# Tests if processing a message returns attention weights as numpy array.
diagnostic_data = processed_message.get(DIAGNOSTIC_DATA)
# DIETClassifier should add attention weights
name = "DIETClassifier_node_name"
assert isinstance(diagnostic_data, dict)
assert name in diagnostic_data
assert "attention_weights" in diagnostic_data[name]
assert isinstance(diagnostic_data[name].get("attention_weights"),
np.ndarray)
assert "text_transformed" in diagnostic_data[name]
assert isinstance(diagnostic_data[name].get("text_transformed"),
np.ndarray)
else:
assert DIAGNOSTIC_DATA not in processed_message.data
async def test_process_gives_diagnostic_data(
default_execution_context: ExecutionContext,
create_response_selector: Callable[[Dict[Text, Any]], ResponseSelector],
train_and_preprocess: Callable[..., Tuple[TrainingData,
List[GraphComponent]]],
process_message: Callable[..., Message],
):
"""Tests if processing a message returns attention weights as numpy array."""
pipeline = [
{
"component": WhitespaceTokenizer
},
{
"component": CountVectorsFeaturizer
},
]
config_params = {EPOCHS: 1}
importer = RasaFileImporter(
config_file="data/test_response_selector_bot/config.yml",
domain_path="data/test_response_selector_bot/domain.yml",
training_data_paths=[
"data/test_response_selector_bot/data/rules.yml",
"data/test_response_selector_bot/data/stories.yml",
"data/test_response_selector_bot/data/nlu.yml",
],
)
training_data = importer.get_nlu_data()
training_data, loaded_pipeline = train_and_preprocess(
pipeline, training_data)
default_execution_context.should_add_diagnostic_data = True
response_selector = create_response_selector(config_params)
response_selector.train(training_data=training_data)
message = Message(data={TEXT: "hello"})
message = process_message(loaded_pipeline, message)
classified_message = response_selector.process([message])[0]
diagnostic_data = classified_message.get(DIAGNOSTIC_DATA)
assert isinstance(diagnostic_data, dict)
for _, values in diagnostic_data.items():
assert "text_transformed" in values
assert isinstance(values.get("text_transformed"), np.ndarray)
# The `attention_weights` key should exist, regardless of there
# being a transformer
assert "attention_weights" in values
# By default, ResponseSelector has `number_of_transformer_layers = 0`
# in which case the attention weights should be None.
assert values.get("attention_weights") is None