def __init__(self, entity_extractor=None, intent_classifier=None, language_name='en', **kwargs):
self.nlp = spacy.load(language_name, parser=False, entity=False, matcher=False)
self.featurizer = SpacyFeaturizer(self.nlp)
with open(intent_classifier, 'rb') as f:
self.classifier = cloudpickle.load(f)
self.extractor = SpacyEntityExtractor(self.nlp, entity_extractor)
def test_spacy_featurizer(spacy_nlp_en, sentence, expected):
from rasa_nlu.featurizers.spacy_featurizer import SpacyFeaturizer
ftr = SpacyFeaturizer(spacy_nlp_en)
doc = spacy_nlp_en(sentence)
vecs = ftr.features_for_doc(doc)
assert np.allclose(doc.vector[:5], expected, atol=1e-5)
assert np.allclose(vecs, doc.vector, atol=1e-5)
def test_spacy_featurizer(sentence, expected, spacy_nlp):
from rasa_nlu.featurizers.spacy_featurizer import SpacyFeaturizer
ftr = SpacyFeaturizer()
doc = spacy_nlp(sentence)
vecs = ftr.features_for_doc(doc)
assert np.allclose(doc.vector[:5], expected, atol=1e-5)
assert np.allclose(vecs, doc.vector, atol=1e-5)
def test_spacy_featurizer(sentence, language, expected):
import spacy
from rasa_nlu.featurizers.spacy_featurizer import SpacyFeaturizer
nlp = spacy.load(language, tagger=False, parser=False)
ftr = SpacyFeaturizer(nlp)
doc = nlp(sentence)
vecs = ftr.features_for_doc(doc)
assert np.allclose(doc.vector[:5], expected, atol=1e-5)
assert np.allclose(vecs, doc.vector, atol=1e-5)
def test_sentence(sentence, language, _ref):
import spacy
from rasa_nlu.featurizers.spacy_featurizer import SpacyFeaturizer
nlp = spacy.load(language, tagger=False, parser=False)
doc = nlp(sentence)
ftr = SpacyFeaturizer(nlp)
vecs = ftr.create_bow_vecs([sentence])
assert np.allclose(doc.vector[:5], _ref, atol=1e-5)
assert np.allclose(vecs[0], doc.vector, atol=1e-5)
def __init__(self, config, language_name):
self.ensure_language_support(language_name)
self.name = "spacy_sklearn"
self.language_name = language_name
self.training_data = None
self.nlp = spacy.load(self.language_name, parser=False, entity=False)
self.featurizer = SpacyFeaturizer(self.nlp)
self.intent_classifier = None
self.entity_extractor = None
def __init__(self,
intent_classifier=None,
entity_extractor=None,
entity_synonyms=None,
nlp=None):
self.extractor = entity_extractor
self.classifier = intent_classifier
self.ent_synonyms = entity_synonyms
self.nlp = nlp
self.featurizer = SpacyFeaturizer(nlp)
ensure_proper_language_model(nlp)
class SpacySklearnInterpreter(Interpreter):
def __init__(self, entity_extractor=None, intent_classifier=None, language_name='en', **kwargs):
self.nlp = spacy.load(language_name, parser=False, entity=False, matcher=False)
self.featurizer = SpacyFeaturizer(self.nlp)
with open(intent_classifier, 'rb') as f:
self.classifier = cloudpickle.load(f)
self.extractor = SpacyEntityExtractor(self.nlp, entity_extractor)
def get_intent(self, text):
"""Returns the most likely intent and its probability for the input text.
:param text: text to classify
:return: tuple of most likely intent name and its probability"""
X = self.featurizer.create_bow_vecs([text])
intent_ids, probabilities = self.classifier.predict(X)
intents = self.classifier.transform_labels_num2str(intent_ids)
return intents[0], probabilities[0]
def parse(self, text):
"""Parse the input text, classify it and return an object containing its intent and entities."""
intent, probability = self.get_intent(text)
entities = self.extractor.extract_entities(self.nlp, text)
return {'text': text, 'intent': intent, 'entities': entities, 'confidence': probability}
def load(meta, nlp, featurizer=None):
"""
:type meta: rasa_nlu.model.Metadata
:type nlp: spacy.language.Language
:type featurizer: None or rasa_nlu.featurizers.spacy_featurizer.SpacyFeaturizer
:rtype: MITIEInterpreter
"""
if meta.entity_extractor_path:
extractor = SpacyEntityExtractor(nlp, meta.entity_extractor_path,
meta.metadata.get("should_fine_tune_spacy_ner"))
else:
extractor = None
if meta.intent_classifier_path:
with open(meta.intent_classifier_path, 'rb') as f:
classifier = cloudpickle.load(f)
else:
classifier = None
if meta.entity_synonyms_path:
entity_synonyms = Interpreter.load_synonyms(meta.entity_synonyms_path)
else:
entity_synonyms = None
if featurizer is None:
featurizer = SpacyFeaturizer(nlp)
return SpacySklearnInterpreter(
classifier,
extractor,
entity_synonyms,
featurizer,
nlp)
class SpacySklearnInterpreter(Interpreter):
def __init__(self,
entity_extractor=None,
intent_classifier=None,
language_name='en',
**kwargs):
self.nlp = spacy.load(language_name,
parser=False,
entity=False,
matcher=False)
self.featurizer = SpacyFeaturizer(self.nlp)
with open(intent_classifier, 'rb') as f:
self.classifier = cloudpickle.load(f)
self.extractor = SpacyEntityExtractor(self.nlp, entity_extractor)
def get_intent(self, text):
X = self.featurizer.create_bow_vecs([text])
return self.classifier.predict(X)[0]
def parse(self, text):
intent = self.get_intent(text)
entities = self.extractor.extract_entities(self.nlp, text)
return {'text': text, 'intent': intent, 'entities': entities}
class SpacySklearnInterpreter(Interpreter):
def __init__(self,
entity_extractor=None,
entity_synonyms=None,
intent_classifier=None,
language_name='en',
**kwargs):
self.extractor = None
self.classifier = None
self.ent_synonyms = None
self.nlp = spacy.load(language_name,
parser=False,
entity=False,
matcher=False)
self.featurizer = SpacyFeaturizer(self.nlp)
ensure_proper_language_model(self.nlp)
if intent_classifier:
with open(intent_classifier, 'rb') as f:
self.classifier = cloudpickle.load(f)
if entity_extractor:
self.extractor = SpacyEntityExtractor(self.nlp, entity_extractor)
self.ent_synonyms = Interpreter.load_synonyms(entity_synonyms)
def get_intent(self, doc):
"""Returns the most likely intent and its probability for the input text.
:param text: text to classify
:return: tuple of most likely intent name and its probability"""
if self.classifier:
X = self.featurizer.features_for_doc(doc).reshape(1, -1)
intent_ids, probabilities = self.classifier.predict(X)
intents = self.classifier.transform_labels_num2str(intent_ids)
intent, score = intents[0], probabilities[0]
else:
intent, score = "None", 0.0
return intent, score
def get_entities(self, doc):
if self.extractor:
return self.extractor.extract_entities(doc)
return []
def parse(self, text):
"""Parse the input text, classify it and return an object containing its intent and entities."""
doc = self.nlp(text)
intent, probability = self.get_intent(doc)
entities = self.get_entities(doc)
if self.ent_synonyms:
Interpreter.replace_synonyms(entities, self.ent_synonyms)
return {
'text': text,
'intent': intent,
'entities': entities,
'confidence': probability
}
class SpacySklearnTrainer(Trainer):
SUPPORTED_LANGUAGES = {"en", "de"}
def __init__(self, config, language_name):
self.ensure_language_support(language_name)
self.name = "spacy_sklearn"
self.language_name = language_name
self.training_data = None
self.nlp = spacy.load(self.language_name, parser=False, entity=False)
self.featurizer = SpacyFeaturizer(self.nlp)
self.intent_classifier = SklearnIntentClassifier()
self.entity_extractor = SpacyEntityExtractor()
def train(self, data, test_split_size=0.1):
self.training_data = data
self.train_entity_extractor(data.entity_examples)
self.train_intent_classifier(data.intent_examples, test_split_size)
def train_entity_extractor(self, entity_examples):
self.entity_extractor.train(self.nlp, entity_examples)
def train_intent_classifier(self, intent_examples, test_split_size=0.1):
labels = [e["intent"] for e in intent_examples]
sentences = [e["text"] for e in intent_examples]
y = self.intent_classifier.transform_labels_str2num(labels)
X = self.featurizer.create_bow_vecs(sentences)
self.intent_classifier.train(X, y, test_split_size)
def persist(self, path, persistor=None, create_unique_subfolder=True):
timestamp = datetime.datetime.now().strftime('%Y%m%d-%H%M%S')
if create_unique_subfolder:
dir_name = os.path.join(path, "model_" + timestamp)
os.mkdir(dir_name)
else:
dir_name = path
data_file = os.path.join(dir_name, "training_data.json")
classifier_file = os.path.join(dir_name, "intent_classifier.pkl")
ner_dir = os.path.join(dir_name, 'ner')
if not os.path.exists(ner_dir):
os.mkdir(ner_dir)
entity_extractor_config_file = os.path.join(ner_dir, "config.json")
entity_extractor_file = os.path.join(ner_dir, "model")
write_training_metadata(dir_name, timestamp, data_file, self.name,
self.language_name, classifier_file, ner_dir)
with open(data_file, 'w') as f:
f.write(self.training_data.as_json(indent=2))
with open(classifier_file, 'wb') as f:
cloudpickle.dump(self.intent_classifier, f)
with open(entity_extractor_config_file, 'w') as f:
json.dump(self.entity_extractor.ner.cfg, f)
self.entity_extractor.ner.model.dump(entity_extractor_file)
if persistor is not None:
persistor.send_tar_to_s3(dir_name)
def __init__(self,
language_name,
max_num_threads=1,
should_fine_tune_spacy_ner=False):
super(self.__class__, self).__init__(language_name, max_num_threads)
self.should_fine_tune_spacy_ner = should_fine_tune_spacy_ner
self.nlp = self._load_nlp_model(language_name,
should_fine_tune_spacy_ner)
self.featurizer = SpacyFeaturizer(self.nlp)
ensure_proper_language_model(self.nlp)
def __init__(self,
entity_extractor=None,
entity_synonyms=None,
intent_classifier=None,
language_name='en',
**kwargs):
self.extractor = None
self.classifier = None
self.ent_synonyms = None
self.nlp = spacy.load(language_name,
parser=False,
entity=False,
matcher=False)
self.featurizer = SpacyFeaturizer(self.nlp)
ensure_proper_language_model(self.nlp)
if intent_classifier:
with open(intent_classifier, 'rb') as f:
self.classifier = cloudpickle.load(f)
if entity_extractor:
self.extractor = SpacyEntityExtractor(self.nlp, entity_extractor)
self.ent_synonyms = Interpreter.load_synonyms(entity_synonyms)
def __init__(self, language_name, max_num_threads=1):
super(self.__class__, self).__init__("spacy_sklearn", language_name,
max_num_threads)
self.nlp = spacy.load(self.language_name, parser=False, entity=False)
self.featurizer = SpacyFeaturizer(self.nlp)
ensure_proper_language_model(self.nlp)
class SpacySklearnInterpreter(Interpreter):
@staticmethod
def load(meta, nlp):
"""
:type meta: ModelMetadata
:rtype: MITIEInterpreter
"""
if meta.entity_extractor_path:
extractor = SpacyEntityExtractor(nlp, meta.entity_extractor_path)
else:
extractor = None
if meta.intent_classifier_path:
with open(meta.intent_classifier_path, 'rb') as f:
classifier = cloudpickle.load(f)
else:
classifier = None
if meta.entity_synonyms_path:
entity_synonyms = Interpreter.load_synonyms(
meta.entity_synonyms_path)
else:
entity_synonyms = None
return SpacySklearnInterpreter(classifier, extractor, entity_synonyms,
nlp)
def __init__(self,
intent_classifier=None,
entity_extractor=None,
entity_synonyms=None,
nlp=None):
self.extractor = entity_extractor
self.classifier = intent_classifier
self.ent_synonyms = entity_synonyms
self.nlp = nlp
self.featurizer = SpacyFeaturizer(nlp)
ensure_proper_language_model(nlp)
def get_intent(self, doc):
"""Returns the most likely intent and its probability for the input text.
:param text: text to classify
:return: tuple of most likely intent name and its probability"""
if self.classifier:
X = self.featurizer.features_for_doc(doc).reshape(1, -1)
intent_ids, probabilities = self.classifier.predict(X)
intents = self.classifier.transform_labels_num2str(intent_ids)
intent, score = intents[0], probabilities[0]
else:
intent, score = "None", 0.0
return intent, score
def get_entities(self, doc):
if self.extractor:
return self.extractor.extract_entities(doc)
return []
def parse(self, text):
"""Parse the input text, classify it and return an object containing its intent and entities."""
doc = self.nlp(text)
intent, probability = self.get_intent(doc)
entities = self.get_entities(doc)
if self.ent_synonyms:
Interpreter.replace_synonyms(entities, self.ent_synonyms)
return {
'text': text,
'intent': intent,
'entities': entities,
'confidence': probability
}
from rasa_nlu.train import load_data
from rasa_nlu.config import RasaNLUModelConfig
from rasa_nlu.utils.spacy_utils import SpacyNLP
from rasa_nlu.tokenizers.spacy_tokenizer import SpacyTokenizer
from rasa_nlu.featurizers.spacy_featurizer import SpacyFeaturizer
import numpy as np, spacy
training_data = load_data("data/examples/rasa/demo-rasa.json")
config = RasaNLUModelConfig()
SpacyNLP(nlp=spacy.load("en")).train(training_data, config)
SpacyTokenizer().train(training_data, config)
SpacyFeaturizer().train(training_data, config)
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import GridSearchCV
from sklearn.svm import SVC
labels = [e.get("intent") for e in training_data.intent_examples]
le = LabelEncoder()
y = le.fit_transform(labels)
X = np.stack([
example.get("text_features") for example in training_data.intent_examples
])
defaults = {
# C parameter of the svm - cross validation will select the best value
"C": [1, 2, 5, 10, 20, 100],
# the kernels to use for the svm training - cross validation will
# decide which one of them performs best