async def test_train_persist_load_with_composite_entities(
component_builder: ComponentBuilder, tmp_path: Path):
pipeline = pipeline_from_components("WhitespaceTokenizer",
"CRFEntityExtractor")
_config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
(trainer, trained, persisted_path) = await rasa.nlu.train.train(
_config,
path=str(tmp_path),
data="data/test/demo-rasa-composite-entities.yml",
component_builder=component_builder,
)
assert trainer.pipeline
assert trained.pipeline
loaded = Interpreter.load(persisted_path, component_builder)
assert loaded.pipeline
text = "I am looking for an italian restaurant"
assert loaded.parse(text) == trained.parse(text)
def test_spacy_featurizer_casing(spacy_nlp):
from rasa.nlu.featurizers.dense_featurizer.spacy_featurizer import SpacyFeaturizer
# if this starts failing for the default model, we should think about
# removing the lower casing the spacy nlp component does when it
# retrieves vectors. For compressed spacy models (e.g. models
# ending in _sm) this test will most likely fail.
ftr = SpacyFeaturizer.create({}, RasaNLUModelConfig())
td = training_data.load_data("data/examples/rasa/demo-rasa.json")
for e in td.intent_examples:
doc = spacy_nlp(e.text)
doc_capitalized = spacy_nlp(e.text.capitalize())
vecs = ftr._features_for_doc(doc)
vecs_capitalized = ftr._features_for_doc(doc_capitalized)
assert np.allclose(
vecs, vecs_capitalized,
atol=1e-5), "Vectors are unequal for texts '{}' and '{}'".format(
e.text, e.text.capitalize())
async def test_train_tensorboard_logging(component_builder, tmpdir):
from pathlib import Path
tensorboard_log_dir = Path(tmpdir.strpath) / "tensorboard"
assert not tensorboard_log_dir.exists()
_config = RasaNLUModelConfig({
"pipeline": [
{
"name": "WhitespaceTokenizer"
},
{
"name": "CountVectorsFeaturizer"
},
{
"name": "DIETClassifier",
EPOCHS: 3,
TENSORBOARD_LOG_LEVEL: "epoch",
TENSORBOARD_LOG_DIR: str(tensorboard_log_dir),
EVAL_NUM_EXAMPLES: 15,
EVAL_NUM_EPOCHS: 1,
},
],
"language":
"en",
})
await train(
_config,
path=tmpdir.strpath,
data="data/examples/rasa/demo-rasa-multi-intent.md",
component_builder=component_builder,
)
assert tensorboard_log_dir.exists()
all_files = list(tensorboard_log_dir.rglob("*.*"))
assert len(all_files) == 3
async def test_inner_linear_normalization(
component_builder: ComponentBuilder,
tmp_path: Path,
classifier_params: Dict[Text, Any],
data_path: Text,
monkeypatch: MonkeyPatch,
):
pipeline = as_pipeline("WhitespaceTokenizer", "CountVectorsFeaturizer",
"DIETClassifier")
assert pipeline[2]["name"] == "DIETClassifier"
pipeline[2].update(classifier_params)
_config = RasaNLUModelConfig({"pipeline": pipeline})
(trained_model, _, persisted_path) = await rasa.nlu.train.train(
_config,
path=str(tmp_path),
data=data_path,
component_builder=component_builder,
)
loaded = Interpreter.load(persisted_path, component_builder)
mock = Mock()
monkeypatch.setattr(train_utils, "normalize", mock.normalize)
parse_data = loaded.parse("hello")
intent_ranking = parse_data.get("intent_ranking")
# check whether normalization had the expected effect
output_sums_to_1 = sum([
intent.get("confidence") for intent in intent_ranking
]) == pytest.approx(1)
assert output_sums_to_1
# check whether the normalization of rankings is reflected in intent prediction
assert parse_data.get("intent") == intent_ranking[0]
# normalize shouldn't have been called
mock.normalize.assert_not_called()
def test_train_selector(pipeline, component_builder, tmpdir):
# use data that include some responses
training_data = rasa.shared.nlu.training_data.loading.load_data(
"data/examples/rasa/demo-rasa.md")
training_data_responses = rasa.shared.nlu.training_data.loading.load_data(
"data/examples/rasa/demo-rasa-responses.md")
training_data = training_data.merge(training_data_responses)
nlu_config = RasaNLUModelConfig({"language": "en", "pipeline": pipeline})
trainer = Trainer(nlu_config)
trainer.train(training_data)
persisted_path = trainer.persist(tmpdir)
assert trainer.pipeline
loaded = Interpreter.load(persisted_path, component_builder)
parsed = loaded.parse("hello")
assert loaded.pipeline
assert parsed is not None
assert (parsed.get("response_selector").get("all_retrieval_intents")) == [
"chitchat"
]
assert (parsed.get("response_selector").get("default").get("response").get(
"intent_response_key")) is not None
assert (parsed.get("response_selector").get("default").get("response").get(
"template_name")) is not None
assert (parsed.get("response_selector").get("default").get("response").get(
"response_templates")) is not None
ranking = parsed.get("response_selector").get("default").get("ranking")
assert ranking is not None
for rank in ranking:
assert rank.get("confidence") is not None
assert rank.get("intent_response_key") is not None
async def test_raise_error_on_incorrect_pipeline(component_builder, tmp_path: Path):
_config = RasaNLUModelConfig(
{
"pipeline": [
{"name": "WhitespaceTokenizer"},
{"name": "DIETClassifier", EPOCHS: 1},
],
"language": "en",
}
)
with pytest.raises(Exception) as e:
await train(
_config,
path=str(tmp_path),
data=DEFAULT_DATA_PATH,
component_builder=component_builder,
)
assert (
"'DIETClassifier' requires ['Featurizer']. "
"Add required components to the pipeline." in str(e.value)
)
def test_train_model_without_data():
td = load_data(DEFAULT_DATA_PATH)
# language, pipeline = pipelines_for_tests()[1]
# show_dict(pipeline)
# exit()
language = "en"
pipeline = load_json(
"{}/test_case/test_pipelines/config_pipeline.json".format(prj_dir))
# exit()
_config = RasaNLUModelConfig({"pipeline": pipeline, "language": language})
trainer = Trainer(_config)
trainer.train(td)
persisted_path = trainer.persist(model_dir)
loaded = Interpreter.load(persisted_path)
assert loaded.pipeline
# Inference
# result = loaded.parse("i'm looking for a place in the north of town")
result = loaded.parse("show me chinese restaurants")
result = dict(
filter(lambda item: item[0] not in ["intent_ranking"], result.items()))
show_dict(result)
async def test_train_model_training_data_persisted(component_builder, tmpdir):
_config = RasaNLUModelConfig({
"pipeline": [{
"name": "KeywordIntentClassifier"
}],
"language":
"en"
})
(trained, _, persisted_path) = await train(
_config,
path=tmpdir.strpath,
data=DEFAULT_DATA_PATH,
component_builder=component_builder,
persist_nlu_training_data=True,
)
assert trained.pipeline
loaded = Interpreter.load(persisted_path, component_builder)
assert loaded.pipeline
assert loaded.model_metadata.get("training_data") is not None
async def test_elmo_train(component_builder, tmpdir):
pipeline = [
{"name": "WhitespaceTokenizer"},
{"name": "ElmoFeaturizer"},
{"name": "CountVectorsFeaturizer"},
{"name": "EmbeddingIntentClassifier"},
]
_config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
(trained, _, persisted_path) = await train(
_config,
path=tmpdir.strpath,
data=DEFAULT_DATA_PATH,
component_builder=component_builder,
)
assert trained.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
def test_train_selector(pipeline, component_builder, tmpdir):
# use data that include some responses
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 = RasaNLUModelConfig({"language": "en", "pipeline": pipeline})
trainer = Trainer(nlu_config)
trainer.train(td)
persisted_path = trainer.persist(tmpdir)
assert trainer.pipeline
loaded = Interpreter.load(persisted_path, component_builder)
parsed = loaded.parse("hello")
assert loaded.pipeline
assert parsed is not None
assert (parsed.get(RESPONSE_SELECTOR_PROPERTY_NAME).get("default").get(
"full_retrieval_intent")) is not None
def test_convert_featurizer_train(component_builder):
tokenizer = component_builder.create_component_from_class(ConveRTTokenizer)
featurizer = component_builder.create_component_from_class(
ConveRTFeaturizer)
sentence = "Hey how are you today ?"
message = Message(sentence)
message.set(RESPONSE, sentence)
tokens = tokenizer.tokenize(message, attribute=TEXT)
tokens = tokenizer.add_cls_token(tokens, attribute=TEXT)
message.set(TOKENS_NAMES[TEXT], tokens)
message.set(TOKENS_NAMES[RESPONSE], tokens)
featurizer.train(TrainingData([message]),
RasaNLUModelConfig(),
tf_hub_module=tokenizer.module)
expected = np.array(
[2.2636216, -0.26475656, -1.1358104, -0.49751878, -1.3946456])
expected_cls = np.array(
[1.0251294, -0.04053932, -0.7018805, -0.82054937, -0.75054353])
vecs = message.get_dense_features(TEXT, [])
assert len(tokens) == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
vecs = message.get_dense_features(RESPONSE, [])
assert len(tokens) == len(vecs)
assert np.allclose(vecs[0][:5], expected, atol=1e-5)
assert np.allclose(vecs[-1][:5], expected_cls, atol=1e-5)
vecs = message.get_dense_features(INTENT, [])
assert vecs is None
def train(self, request):
# if not set will use the default project name, e.g. "default"
project = parameter_or_default(request, "project", default=None)
# if set will not generate a model name but use the passed one
model_name = parameter_or_default(request, "model", default=None)
try:
model_config, data = self.extract_data_and_config(request)
except Exception as e:
request.setResponseCode(400)
returnValue(json_to_string({"error": "{}".format(e)}))
data_file = dump_to_data_file(data)
request.setHeader('Content-Type', 'application/zip')
try:
request.setResponseCode(200)
request.setHeader("Content-Disposition", "attachment")
path_to_model = yield self.data_router.start_train_process(
data_file, project, RasaNLUModelConfig(model_config),
model_name)
zipped_path = utils.zip_folder(path_to_model)
zip_content = io.open(zipped_path, 'r+b').read()
return returnValue(zip_content)
except MaxTrainingError as e:
request.setResponseCode(403)
returnValue(json_to_string({"error": "{}".format(e)}))
except InvalidProjectError as e:
request.setResponseCode(404)
returnValue(json_to_string({"error": "{}".format(e)}))
except TrainingException as e:
request.setResponseCode(500)
returnValue(json_to_string({"error": "{}".format(e)}))
async def test_softmax_normalization(
component_builder,
tmp_path,
classifier_params,
data_path,
output_length,
output_should_sum_to_1,
):
pipeline = as_pipeline(
"WhitespaceTokenizer", "CountVectorsFeaturizer", "DIETClassifier"
)
assert pipeline[2]["name"] == "DIETClassifier"
pipeline[2].update(classifier_params)
_config = RasaNLUModelConfig({"pipeline": pipeline})
(trained_model, _, persisted_path) = await train(
_config,
path=str(tmp_path),
data=data_path,
component_builder=component_builder,
)
loaded = Interpreter.load(persisted_path, component_builder)
parse_data = loaded.parse("hello")
intent_ranking = parse_data.get("intent_ranking")
# check that the output was correctly truncated after normalization
assert len(intent_ranking) == output_length
# check whether normalization had the expected effect
output_sums_to_1 = sum(
[intent.get("confidence") for intent in intent_ranking]
) == pytest.approx(1)
assert output_sums_to_1 == output_should_sum_to_1
# check whether the normalization of rankings is reflected in intent prediction
assert parse_data.get("intent") == intent_ranking[0]
async def test_set_random_seed(component_builder, tmpdir):
"""test if train result is the same for two runs of tf embedding"""
# set fixed random seed
_config = RasaNLUModelConfig(
{
"pipeline": [
{"name": "WhitespaceTokenizer"},
{"name": "CountVectorsFeaturizer"},
{"name": "DIETClassifier", RANDOM_SEED: 1, EPOCHS: 1},
],
"language": "en",
}
)
# first run
(trained_a, _, persisted_path_a) = await train(
_config,
path=tmpdir.strpath + "_a",
data=DEFAULT_DATA_PATH,
component_builder=component_builder,
)
# second run
(trained_b, _, persisted_path_b) = await train(
_config,
path=tmpdir.strpath + "_b",
data=DEFAULT_DATA_PATH,
component_builder=component_builder,
)
loaded_a = Interpreter.load(persisted_path_a, component_builder)
loaded_b = Interpreter.load(persisted_path_b, component_builder)
result_a = loaded_a.parse("hello")["intent"]["confidence"]
result_b = loaded_b.parse("hello")["intent"]["confidence"]
assert result_a == result_b
def test_spacy_featurizer_using_empty_model():
from rasa.nlu.featurizers.dense_featurizer.spacy_featurizer import SpacyFeaturizer
import spacy
sentence = "This test is using an empty spaCy model"
model = spacy.blank("en")
doc = model(sentence)
ftr = SpacyFeaturizer.create({}, RasaNLUModelConfig())
message = Message(data={TEXT: sentence})
message.set(SPACY_DOCS[TEXT], doc)
ftr._set_spacy_features(message)
seq_vecs, sen_vecs = message.get_dense_features(TEXT, [])
if seq_vecs:
seq_vecs = seq_vecs.features
if sen_vecs:
sen_vecs = sen_vecs.features
assert seq_vecs is None
assert sen_vecs is None
async def test_train_persist_load_with_composite_entities(
classifier_params, component_builder, tmpdir):
pipeline = as_pipeline("WhitespaceTokenizer", "CountVectorsFeaturizer",
"DIETClassifier")
assert pipeline[2]["name"] == "DIETClassifier"
pipeline[2].update(classifier_params)
_config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
(trainer, trained, persisted_path) = await rasa.nlu.train.train(
_config,
path=tmpdir.strpath,
data="data/test/demo-rasa-composite-entities.yml",
component_builder=component_builder,
)
assert trainer.pipeline
assert trained.pipeline
loaded = Interpreter.load(persisted_path, component_builder)
assert loaded.pipeline
text = "I am looking for an italian restaurant"
assert loaded.parse(text) == trained.parse(text)
def test_duckling_entity_extractor_and_synonyms(component_builder):
_config = RasaNLUModelConfig({
"pipeline": [{
"name": "DucklingHTTPExtractor"
}, {
"name": "EntitySynonymMapper"
}]
})
_config.set_component_attr(0, dimensions=["number"])
duckling = component_builder.create_component(_config.for_component(0),
_config)
synonyms = component_builder.create_component(_config.for_component(1),
_config)
message = Message("He was 6 feet away")
duckling.process(message)
# checks that the synonym processor
# can handle entities that have int values
synonyms.process(message)
assert message is not None
def test_set_attr_on_component():
_config = RasaNLUModelConfig(
{
"language": "en",
"pipeline": [
{"name": "SpacyNLP"},
{"name": "SpacyTokenizer"},
{"name": "SpacyFeaturizer"},
{"name": "DIETClassifier"},
],
}
)
idx_classifier = _config.component_names.index("DIETClassifier")
idx_tokenizer = _config.component_names.index("SpacyTokenizer")
_config.set_component_attr(idx_classifier, epochs=10)
assert _config.for_component(idx_tokenizer) == {"name": "SpacyTokenizer"}
assert _config.for_component(idx_classifier) == {
"name": "DIETClassifier",
"epochs": 10,
}
def blank_config() -> RasaNLUModelConfig:
return RasaNLUModelConfig({"language": "en", "pipeline": []})
def test_crf_extractor(spacy_nlp, ner_crf_pos_feature_config):
from rasa.nlu.extractors.crf_entity_extractor import CRFEntityExtractor
ext = CRFEntityExtractor(component_config=ner_crf_pos_feature_config)
examples = [
Message(
"anywhere in the west",
{
"intent":
"restaurant_search",
"entities": [{
"start": 16,
"end": 20,
"value": "west",
"entity": "location"
}],
"spacy_doc":
spacy_nlp("anywhere in the west"),
},
),
Message(
"central indian restaurant",
{
"intent":
"restaurant_search",
"entities": [
{
"start": 0,
"end": 7,
"value": "central",
"entity": "location",
"extractor": "random_extractor",
},
{
"start": 8,
"end": 14,
"value": "indian",
"entity": "cuisine",
"extractor": "CRFEntityExtractor",
},
],
"spacy_doc":
spacy_nlp("central indian restaurant"),
},
),
]
# uses BILOU and the default features
ext.train(TrainingData(training_examples=examples), RasaNLUModelConfig())
sentence = "anywhere in the west"
doc = {"spacy_doc": spacy_nlp(sentence)}
crf_format = ext._from_text_to_crf(Message(sentence, doc))
assert [word[0]
for word in crf_format] == ["anywhere", "in", "the", "west"]
feats = ext._sentence_to_features(crf_format)
assert "BOS" in feats[0]
assert "EOS" in feats[-1]
assert feats[1]["0:low"] == "in"
sentence = "anywhere in the west"
ext.extract_entities(Message(sentence, {"spacy_doc": spacy_nlp(sentence)}))
filtered = ext.filter_trainable_entities(examples)
assert filtered[0].get("entities") == [{
"start": 16,
"end": 20,
"value": "west",
"entity": "location"
}], "Entity without extractor remains"
assert filtered[1].get("entities") == [{
"start": 8,
"end": 14,
"value": "indian",
"entity": "cuisine",
"extractor": "CRFEntityExtractor",
}], "Only CRFEntityExtractor entity annotation remains"
assert examples[1].get("entities")[0] == {
"start": 0,
"end": 7,
"value": "central",
"entity": "location",
"extractor": "random_extractor",
}, "Original examples are not mutated"
def test_duckling_entity_extractor(component_builder):
httpretty.register_uri(
httpretty.POST,
"http://localhost:8000/parse",
body="""[{"body":"Today","start":0,"value":{"values":[{
"value":"2018-11-13T00:00:00.000-08:00","grain":"day",
"type":"value"}],"value":"2018-11-13T00:00:00.000-08:00",
"grain":"day","type":"value"},"end":5,
"dim":"time","latent":false},{"body":"the 5th","start":9,
"value":{"values":[{
"value":"2018-12-05T00:00:00.000-08:00","grain":"day",
"type":"value"},
{"value":"2019-01-05T00:00:00.000-08:00","grain":"day",
"type":"value"},
{"value":"2019-02-05T00:00:00.000-08:00","grain":"day",
"type":"value"}],
"value":"2018-12-05T00:00:00.000-08:00","grain":"day",
"type":"value"},"end":16,"dim":"time",
"latent":false},{"body":"5th of May","start":13,"value":{
"values":[{
"value":"2019-05-05T00:00:00.000-07:00","grain":"day",
"type":"value"},
{"value":"2020-05-05T00:00:00.000-07:00","grain":"day",
"type":"value"},
{"value":"2021-05-05T00:00:00.000-07:00","grain":"day",
"type":"value"}],
"value":"2019-05-05T00:00:00.000-07:00","grain":"day",
"type":"value"},"end":23,"dim":"time",
"latent":false},{"body":"tomorrow","start":37,"value":{
"values":[{
"value":"2018-11-14T00:00:00.000-08:00","grain":"day",
"type":"value"}],
"value":"2018-11-14T00:00:00.000-08:00","grain":"day",
"type":"value"},"end":45,"dim":"time",
"latent":false}]""",
)
httpretty.enable()
_config = RasaNLUModelConfig(
{"pipeline": [{
"name": "DucklingHTTPExtractor"
}]})
_config.set_component_attr(0,
dimensions=["time"],
timezone="UTC",
url="http://localhost:8000")
duckling = component_builder.create_component(_config.for_component(0),
_config)
message = Message("Today is the 5th of May. Let us meet tomorrow.")
duckling.process(message)
entities = message.get("entities")
assert len(entities) == 4
# Test duckling with a defined date
httpretty.register_uri(
httpretty.POST,
"http://localhost:8000/parse",
body="""[{"body":"tomorrow","start":12,"value":{"values":[{
"value":"2013-10-13T00:00:00.000Z","grain":"day",
"type":"value"}],"value":"2013-10-13T00:00:00.000Z",
"grain":"day","type":"value"},"end":20,
"dim":"time","latent":false}]""",
)
# 1381536182 == 2013/10/12 02:03:02
message = Message("Let us meet tomorrow.", time="1381536182")
duckling.process(message)
entities = message.get("entities")
assert len(entities) == 1
assert entities[0]["text"] == "tomorrow"
assert entities[0]["value"] == "2013-10-13T00:00:00.000Z"
# Test dimension filtering includes only specified dimensions
_config = RasaNLUModelConfig(
{"pipeline": [{
"name": "DucklingHTTPExtractor"
}]})
_config.set_component_attr(0,
dimensions=["number"],
url="http://localhost:8000")
ducklingNumber = component_builder.create_component(
_config.for_component(0), _config)
httpretty.register_uri(
httpretty.POST,
"http://localhost:8000/parse",
body="""[{"body":"Yesterday","start":0,"value":{"values":[{
"value":"2019-02-28T00:00:00.000+01:00","grain":"day",
"type":"value"}],"value":"2019-02-28T00:00:00.000+01:00",
"grain":"day","type":"value"},"end":9,"dim":"time"},
{"body":"5","start":21,"value":{"value":5,"type":"value"},
"end":22,"dim":"number"}]""",
)
message = Message("Yesterday there were 5 people in a room")
ducklingNumber.process(message)
entities = message.get("entities")
assert len(entities) == 1
assert entities[0]["text"] == "5"
assert entities[0]["value"] == 5
def test_run_cv_evaluation_with_response_selector(monkeypatch: MonkeyPatch):
training_data_obj = rasa.shared.nlu.training_data.loading.load_data(
"data/examples/rasa/demo-rasa.yml"
)
training_data_responses_obj = rasa.shared.nlu.training_data.loading.load_data(
"data/examples/rasa/demo-rasa-responses.yml"
)
training_data_obj = training_data_obj.merge(training_data_responses_obj)
nlu_config = RasaNLUModelConfig(
{
"language": "en",
"pipeline": [
{"name": "WhitespaceTokenizer"},
{"name": "CountVectorsFeaturizer"},
{"name": "DIETClassifier", EPOCHS: 2},
{"name": "ResponseSelector", EPOCHS: 2},
],
}
)
# mock training
trainer = Trainer(nlu_config)
trainer.pipeline = remove_pretrained_extractors(trainer.pipeline)
mock = Mock(return_value=Interpreter(trainer.pipeline, None))
monkeypatch.setattr(Trainer, "train", mock)
n_folds = 2
intent_results, entity_results, response_selection_results = cross_validate(
training_data_obj,
n_folds,
nlu_config,
successes=False,
errors=False,
disable_plotting=True,
report_as_dict=True,
)
assert len(intent_results.train["Accuracy"]) == n_folds
assert len(intent_results.train["Precision"]) == n_folds
assert len(intent_results.train["F1-score"]) == n_folds
assert len(intent_results.test["Accuracy"]) == n_folds
assert len(intent_results.test["Precision"]) == n_folds
assert len(intent_results.test["F1-score"]) == n_folds
assert all(key in intent_results.evaluation for key in ["errors", "report"])
assert any(
isinstance(intent_report, dict)
and intent_report.get("confused_with") is not None
for intent_report in intent_results.evaluation["report"].values()
)
assert len(response_selection_results.train["Accuracy"]) == n_folds
assert len(response_selection_results.train["Precision"]) == n_folds
assert len(response_selection_results.train["F1-score"]) == n_folds
assert len(response_selection_results.test["Accuracy"]) == n_folds
assert len(response_selection_results.test["Precision"]) == n_folds
assert len(response_selection_results.test["F1-score"]) == n_folds
assert all(
key in response_selection_results.evaluation for key in ["errors", "report"]
)
assert any(
isinstance(intent_report, dict)
and intent_report.get("confused_with") is not None
for intent_report in response_selection_results.evaluation["report"].values()
)
assert len(entity_results.train["DIETClassifier"]["Accuracy"]) == n_folds
assert len(entity_results.train["DIETClassifier"]["Precision"]) == n_folds
assert len(entity_results.train["DIETClassifier"]["F1-score"]) == n_folds
assert len(entity_results.test["DIETClassifier"]["Accuracy"]) == n_folds
assert len(entity_results.test["DIETClassifier"]["Precision"]) == n_folds
assert len(entity_results.test["DIETClassifier"]["F1-score"]) == n_folds
for extractor_evaluation in entity_results.evaluation.values():
assert all(key in extractor_evaluation for key in ["errors", "report"])
def get_nlu_model(self) -> RasaNLUModelConfig:
return RasaNLUModelConfig({
"language": self.language,
"pipeline": self.pipeline
})
def test_run_cv_evaluation_with_response_selector():
training_data_obj = rasa.shared.nlu.training_data.loading.load_data(
"data/examples/rasa/demo-rasa.md")
training_data_responses_obj = rasa.shared.nlu.training_data.loading.load_data(
"data/examples/rasa/demo-rasa-responses.md")
training_data_obj = training_data_obj.merge(training_data_responses_obj)
nlu_config = RasaNLUModelConfig({
"language":
"en",
"pipeline": [
{
"name": "WhitespaceTokenizer"
},
{
"name": "CountVectorsFeaturizer"
},
{
"name": "DIETClassifier",
EPOCHS: 2
},
{
"name": "ResponseSelector",
EPOCHS: 2
},
],
})
n_folds = 2
intent_results, entity_results, response_selection_results = cross_validate(
training_data_obj,
n_folds,
nlu_config,
successes=False,
errors=False,
disable_plotting=True,
)
assert len(intent_results.train["Accuracy"]) == n_folds
assert len(intent_results.train["Precision"]) == n_folds
assert len(intent_results.train["F1-score"]) == n_folds
assert len(intent_results.test["Accuracy"]) == n_folds
assert len(intent_results.test["Precision"]) == n_folds
assert len(intent_results.test["F1-score"]) == n_folds
assert all(key in intent_results.evaluation
for key in ["errors", "report"])
assert len(response_selection_results.train["Accuracy"]) == n_folds
assert len(response_selection_results.train["Precision"]) == n_folds
assert len(response_selection_results.train["F1-score"]) == n_folds
assert len(response_selection_results.test["Accuracy"]) == n_folds
assert len(response_selection_results.test["Precision"]) == n_folds
assert len(response_selection_results.test["F1-score"]) == n_folds
assert all(key in response_selection_results.evaluation
for key in ["errors", "report"])
assert len(entity_results.train["DIETClassifier"]["Accuracy"]) == n_folds
assert len(entity_results.train["DIETClassifier"]["Precision"]) == n_folds
assert len(entity_results.train["DIETClassifier"]["F1-score"]) == n_folds
assert len(entity_results.test["DIETClassifier"]["Accuracy"]) == n_folds
assert len(entity_results.test["DIETClassifier"]["Precision"]) == n_folds
assert len(entity_results.test["DIETClassifier"]["F1-score"]) == n_folds
for extractor_evaluation in entity_results.evaluation.values():
assert all(key in extractor_evaluation for key in ["errors", "report"])
def test_incremental_train_featurization(tmp_path: Path):
patterns = [
{
"pattern": "[0-9]+",
"name": "number",
"usage": "intent"
},
{
"pattern": "\\bhey*",
"name": "hello",
"usage": "intent"
},
{
"pattern": "[0-1]+",
"name": "binary",
"usage": "intent"
},
]
featurizer = RegexFeaturizer.create({"number_additional_patterns": 5},
RasaNLUModelConfig())
sentence = "hey how are you today 19.12.2019 ?"
message = Message(data={TEXT: sentence})
message.set(RESPONSE, sentence)
message.set(INTENT, "intent")
WhitespaceTokenizer().train(TrainingData([message]))
featurizer.train(TrainingData([message], regex_features=patterns),
RasaNLUModelConfig())
# Test featurization of message
expected = np.array([0, 1, 0, 0, 0, 0, 0, 0])
expected_cls = np.array([1, 1, 1, 0, 0, 0, 0, 0])
seq_vecs, sen_vec = message.get_sparse_features(TEXT, [])
if seq_vecs:
seq_vecs = seq_vecs.features
if sen_vec:
sen_vec = sen_vec.features
assert (6, 8) == seq_vecs.shape
assert (1, 8) == sen_vec.shape
assert np.all(seq_vecs.toarray()[0] == expected)
assert np.all(sen_vec.toarray()[-1] == expected_cls)
persist_value = featurizer.persist("ftr", str(tmp_path))
loaded_featurizer = RegexFeaturizer.load(
meta={
"number_additional_patterns": 5,
"file": persist_value["file"],
},
should_finetune=True,
model_dir=str(tmp_path),
)
new_patterns = [
{
"pattern": "\\btoday*",
"name": "day",
"usage": "intent"
},
{
"pattern": "\\bhey+",
"name": "hello",
"usage": "intent"
},
]
message = Message(data={TEXT: sentence})
message.set(RESPONSE, sentence)
message.set(INTENT, "intent")
WhitespaceTokenizer().train(TrainingData([message]))
loaded_featurizer.train(
TrainingData([message], regex_features=patterns + new_patterns),
RasaNLUModelConfig(),
)
# Test featurization of message, this time for the extra pattern as well.
expected_token_1 = np.array([0, 1, 0, 0, 0, 0, 0, 0])
expected_token_2 = np.array([0, 0, 0, 1, 0, 0, 0, 0])
expected_cls = np.array([1, 1, 1, 1, 0, 0, 0, 0])
seq_vecs, sen_vec = message.get_sparse_features(TEXT, [])
if seq_vecs:
seq_vecs = seq_vecs.features
if sen_vec:
sen_vec = sen_vec.features
assert (6, 8) == seq_vecs.shape
assert (1, 8) == sen_vec.shape
assert np.all(seq_vecs.toarray()[0] == expected_token_1)
assert np.all(seq_vecs.toarray()[-2] == expected_token_2)
assert np.all(sen_vec.toarray()[-1] == expected_cls)
# we also modified a pattern, check if that is correctly modified
pattern_to_check = [
pattern for pattern in loaded_featurizer.known_patterns
if pattern["name"] == "hello"
]
assert pattern_to_check == [new_patterns[1]]
def test_persist_load_for_finetuning(tmp_path: Path):
patterns = [
{
"pattern": "[0-9]+",
"name": "number",
"usage": "intent"
},
{
"pattern": "\\bhey*",
"name": "hello",
"usage": "intent"
},
{
"pattern": "[0-1]+",
"name": "binary",
"usage": "intent"
},
]
featurizer = RegexFeaturizer.create({"number_additional_patterns": 5},
RasaNLUModelConfig())
sentence = "hey how are you today 19.12.2019 ?"
message = Message(data={TEXT: sentence})
message.set(RESPONSE, sentence)
message.set(INTENT, "intent")
WhitespaceTokenizer().train(TrainingData([message]))
featurizer.train(TrainingData([message], regex_features=patterns),
RasaNLUModelConfig())
persist_value = featurizer.persist("ftr", str(tmp_path))
# Test all artifacts stored as part of persist
assert persist_value["file"] == "ftr"
assert (tmp_path / "ftr.patterns.pkl").exists()
assert (tmp_path / "ftr.vocabulary_stats.pkl").exists()
assert featurizer.vocabulary_stats == {
"max_number_patterns": 8,
"pattern_slots_filled": 3,
}
loaded_featurizer = RegexFeaturizer.load(
meta={
"number_additional_patterns": 5,
"file": persist_value["file"],
},
should_finetune=True,
model_dir=str(tmp_path),
)
# Test component loaded in finetune mode and also with
# same patterns as before and vocabulary statistics
assert loaded_featurizer.known_patterns == featurizer.known_patterns
assert loaded_featurizer.finetune_mode
assert loaded_featurizer.pattern_vocabulary_stats == featurizer.vocabulary_stats
new_lookups = [{
"name": "plates",
"elements": "data/test/lookup_tables/plates.txt"
}]
training_data = TrainingData()
training_data.lookup_tables = new_lookups
loaded_featurizer.train(training_data)
# Test merging of a new pattern to an already trained component.
assert len(loaded_featurizer.known_patterns) == 4
assert loaded_featurizer.vocabulary_stats == {
"max_number_patterns": 8,
"pattern_slots_filled": 4,
}
def test_regex_featurizer_train():
patterns = [
{
"pattern": "[0-9]+",
"name": "number",
"usage": "intent"
},
{
"pattern": "\\bhey*",
"name": "hello",
"usage": "intent"
},
{
"pattern": "[0-1]+",
"name": "binary",
"usage": "intent"
},
]
featurizer = RegexFeaturizer.create({"number_additional_patterns": 0},
RasaNLUModelConfig())
sentence = "hey how are you today 19.12.2019 ?"
message = Message(data={TEXT: sentence})
message.set(RESPONSE, sentence)
message.set(INTENT, "intent")
WhitespaceTokenizer().train(TrainingData([message]))
featurizer.train(TrainingData([message], regex_features=patterns),
RasaNLUModelConfig())
expected = np.array([0, 1, 0])
expected_cls = np.array([1, 1, 1])
seq_vecs, sen_vec = message.get_sparse_features(TEXT, [])
if seq_vecs:
seq_vecs = seq_vecs.features
if sen_vec:
sen_vec = sen_vec.features
assert (6, 3) == seq_vecs.shape
assert (1, 3) == sen_vec.shape
assert np.all(seq_vecs.toarray()[0] == expected)
assert np.all(sen_vec.toarray()[-1] == expected_cls)
seq_vecs, sen_vec = message.get_sparse_features(RESPONSE, [])
if seq_vecs:
seq_vecs = seq_vecs.features
if sen_vec:
sen_vec = sen_vec.features
assert (6, 3) == seq_vecs.shape
assert (1, 3) == sen_vec.shape
assert np.all(seq_vecs.toarray()[0] == expected)
assert np.all(sen_vec.toarray()[-1] == expected_cls)
seq_vecs, sen_vec = message.get_sparse_features(INTENT, [])
if seq_vecs:
seq_vecs = seq_vecs.features
if sen_vec:
sen_vec = sen_vec.features
assert seq_vecs is None
assert sen_vec is None
async def test_adjusting_layers_incremental_training(
component_builder: ComponentBuilder, tmpdir: Path):
"""Tests adjusting sparse layers of `ResponseSelector` to increased sparse
feature sizes during incremental training.
Testing is done by checking the layer sizes.
Checking if they were replaced correctly is also important
and is done in `test_replace_dense_for_sparse_layers`
in `test_rasa_layers.py`.
"""
iter1_data_path = "data/test_incremental_training/iter1/"
iter2_data_path = "data/test_incremental_training/"
pipeline = [
{
"name": "WhitespaceTokenizer"
},
{
"name": "LexicalSyntacticFeaturizer"
},
{
"name": "RegexFeaturizer"
},
{
"name": "CountVectorsFeaturizer"
},
{
"name": "CountVectorsFeaturizer",
"analyzer": "char_wb",
"min_ngram": 1,
"max_ngram": 4,
},
{
"name": "ResponseSelector",
EPOCHS: 1
},
]
_config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
(_, trained, persisted_path) = await rasa.nlu.train.train(
_config,
path=str(tmpdir),
data=iter1_data_path,
component_builder=component_builder,
)
assert trained.pipeline
old_data_signature = trained.pipeline[-1].model.data_signature
old_predict_data_signature = trained.pipeline[
-1].model.predict_data_signature
message = Message.build(text="Rasa is great!")
trained.featurize_message(message)
old_sparse_feature_sizes = message.get_sparse_feature_sizes(attribute=TEXT)
initial_rs_layers = (
trained.pipeline[-1].model._tf_layers["sequence_layer.text"].
_tf_layers["feature_combining"])
initial_rs_sequence_layer = initial_rs_layers._tf_layers[
"sparse_dense.sequence"]._tf_layers["sparse_to_dense"]
initial_rs_sentence_layer = initial_rs_layers._tf_layers[
"sparse_dense.sentence"]._tf_layers["sparse_to_dense"]
initial_rs_sequence_size = initial_rs_sequence_layer.get_kernel().shape[0]
initial_rs_sentence_size = initial_rs_sentence_layer.get_kernel().shape[0]
assert initial_rs_sequence_size == sum(
old_sparse_feature_sizes[FEATURE_TYPE_SEQUENCE])
assert initial_rs_sentence_size == sum(
old_sparse_feature_sizes[FEATURE_TYPE_SENTENCE])
loaded = Interpreter.load(
persisted_path,
component_builder,
new_config=_config,
)
assert loaded.pipeline
assert loaded.parse("Rasa is great!") == trained.parse("Rasa is great!")
(_, trained, _) = await rasa.nlu.train.train(
_config,
path=str(tmpdir),
data=iter2_data_path,
component_builder=component_builder,
model_to_finetune=loaded,
)
assert trained.pipeline
message = Message.build(text="Rasa is great!")
trained.featurize_message(message)
new_sparse_feature_sizes = message.get_sparse_feature_sizes(attribute=TEXT)
final_rs_layers = (
trained.pipeline[-1].model._tf_layers["sequence_layer.text"].
_tf_layers["feature_combining"])
final_rs_sequence_layer = final_rs_layers._tf_layers[
"sparse_dense.sequence"]._tf_layers["sparse_to_dense"]
final_rs_sentence_layer = final_rs_layers._tf_layers[
"sparse_dense.sentence"]._tf_layers["sparse_to_dense"]
final_rs_sequence_size = final_rs_sequence_layer.get_kernel().shape[0]
final_rs_sentence_size = final_rs_sentence_layer.get_kernel().shape[0]
assert final_rs_sequence_size == sum(
new_sparse_feature_sizes[FEATURE_TYPE_SEQUENCE])
assert final_rs_sentence_size == sum(
new_sparse_feature_sizes[FEATURE_TYPE_SENTENCE])
# check if the data signatures were correctly updated
new_data_signature = trained.pipeline[-1].model.data_signature
new_predict_data_signature = trained.pipeline[
-1].model.predict_data_signature
iter2_data = load_data(iter2_data_path)
expected_sequence_lengths = len([
message for message in iter2_data.training_examples
if message.get(INTENT_RESPONSE_KEY)
])
def test_data_signatures(
new_signature: Dict[Text, Dict[Text, List[FeatureArray]]],
old_signature: Dict[Text, Dict[Text, List[FeatureArray]]],
):
# Wherever attribute / feature_type signature is not
# expected to change, directly compare it to old data signature.
# Else compute its expected signature and compare
attributes_expected_to_change = [TEXT]
feature_types_expected_to_change = [
FEATURE_TYPE_SEQUENCE,
FEATURE_TYPE_SENTENCE,
]
for attribute, signatures in new_signature.items():
for feature_type, feature_signatures in signatures.items():
if feature_type == "sequence_lengths":
assert feature_signatures[
0].units == expected_sequence_lengths
elif feature_type not in feature_types_expected_to_change:
assert feature_signatures == old_signature.get(
attribute).get(feature_type)
else:
for index, feature_signature in enumerate(
feature_signatures):
if (feature_signature.is_sparse and attribute
in attributes_expected_to_change):
assert feature_signature.units == sum(
new_sparse_feature_sizes.get(feature_type))
else:
# dense signature or attributes that are not
# expected to change can be compared directly
assert (
feature_signature.units == old_signature.get(
attribute).get(feature_type)[index].units)
test_data_signatures(new_data_signature, old_data_signature)
test_data_signatures(new_predict_data_signature,
old_predict_data_signature)
def train(self, training_data: TrainingData, config: RasaNLUModelConfig,
**kwargs: Any) -> None:
from seq2label.input import build_input_func
from seq2label.model import Model
raw_config = self.component_config
print(raw_config)
if 'result_dir' not in raw_config:
raw_config['result_dir'] = tempfile.mkdtemp()
model = Model(raw_config)
config = model.get_default_config()
config.update(raw_config)
# task_status = TaskStatus(config)
# read data according configure
train_data_generator_func = kwargs.get('addons_tf_input_fn')
corpus_meta_data = kwargs.get('addons_tf_input_meta')
config['tags_data'] = corpus_meta_data['label']
config['num_classes'] = len(config['tags_data'])
print('')
# build model according configure
# send START status to monitor system
# task_status.send_status(task_status.START)
# train and evaluate model
train_input_func = build_input_func(train_data_generator_func, config)
# train_iterator = train_input_func()
# import tensorflow as tf
# import sys
#
# with tf.Session() as sess:
# sess.run(tf.tables_initializer())
#
# counter = 0
# while True:
# try:
# value = sess.run(train_iterator[0]['words'])
# counter += 1
# print(value)
# break
# except tf.errors.OutOfRangeError:
# break
#
# print(counter)
# #
# sys.exit(0)
evaluate_result, export_results, final_saved_model = model.train_and_eval_then_save(
train_input_func, None, config)
# task_status.send_status(task_status.DONE)
self.result_dir = final_saved_model
async def test_sparse_feature_sizes_decreased_incremental_training(
iter1_path: Text,
iter2_path: Text,
should_raise_exception: bool,
component_builder: ComponentBuilder,
tmpdir: Path,
):
pipeline = [
{
"name": "WhitespaceTokenizer"
},
{
"name": "LexicalSyntacticFeaturizer"
},
{
"name": "RegexFeaturizer"
},
{
"name": "CountVectorsFeaturizer"
},
{
"name": "CountVectorsFeaturizer",
"analyzer": "char_wb",
"min_ngram": 1,
"max_ngram": 4,
},
{
"name": "ResponseSelector",
EPOCHS: 1
},
]
_config = RasaNLUModelConfig({"pipeline": pipeline, "language": "en"})
(_, trained, persisted_path) = await rasa.nlu.train.train(
_config,
path=str(tmpdir),
data=iter1_path,
component_builder=component_builder,
)
assert trained.pipeline
loaded = Interpreter.load(
persisted_path,
component_builder,
new_config=_config,
)
assert loaded.pipeline
assert loaded.parse("Rasa is great!") == trained.parse("Rasa is great!")
if should_raise_exception:
with pytest.raises(Exception) as exec_info:
(_, trained, _) = await rasa.nlu.train.train(
_config,
path=str(tmpdir),
data=iter2_path,
component_builder=component_builder,
model_to_finetune=loaded,
)
assert "Sparse feature sizes have decreased" in str(exec_info.value)
else:
(_, trained, _) = await rasa.nlu.train.train(
_config,
path=str(tmpdir),
data=iter2_path,
component_builder=component_builder,
model_to_finetune=loaded,
)
assert trained.pipeline