def fit(self, x_seq, y_seq, validation_split=0., epochs=1, verbose=0):
with self.graph.as_default():
# Assumption: sequence_tagger(words) -> characters
x_train, lengths = self.lang.tokenized_to_sequence(x_seq)
y_train = utils.one_hot_encode_sequence(y_seq, self.classes)
history = self.model.fit(x_train, y_train, verbose=verbose)
return history.history['loss'][0], history.history['acc'][0]
def fit(self, x_seq, y_seq, validation_split=0., epochs=1):
with self.graph.as_default():
# Assumption: sequence_tagger(words) -> characters
x_train, lengths = self.lang.texts_to_sequence(x_seq)
y_train = utils.one_hot_encode_sequence(
y_seq, valid_tokens=self.valid_outputs)
return self.model.fit(x_train, y_train)
def _score_sequence(self, x_test, y_test):
id2class = self.label_helper.score_classes
self.logger.debug("Scoring items with model labeller.")
y_test = self.label_helper.to_training(y_test)
y_test = utils.one_hot_encode_sequence(
y_test,
id2class,
)
y_pred = self.model.score(x_test)
# reshape as a classification problem, to set threshold
threshold = Evaluator.threshold_for_precision(
y_test.reshape((y_test.shape[0], -1)),
y_pred.reshape((y_test.shape[0], -1)),
TARGET_PRECISION,
)
threshold = 0.45
unlabelled_texts, ids = self.dataset.model_labelling_set()
if len(unlabelled_texts) == 0:
self.logger.info("Model labelling done!")
return 0
scores = self.model.score(unlabelled_texts)
dist = scores / scores.sum(axis=-1, keepdims=True)
idxs = np.argmax(dist, -1)
num_scored = 0
self.logger.debug("set labelling threshold as: {}".format(threshold))
for _id, (text, prediction) in zip(ids, zip(unlabelled_texts, dist)):
# The prediction has padding so we only take the last len(text) scores.
text_tag = []
met_threshold = True
for word, word_likelihood_dist in zip(text,
prediction[-len(text):]):
idx = np.argmax(word_likelihood_dist)
tag = id2class[idx]
if np.max(word_likelihood_dist) < threshold:
met_threshold = False
break
text_tag.append({'word': word, 'tag': tag})
if met_threshold:
self.dataset.add_label(
_id,
self.label_helper.decode(text_tag),
stage=Dataset.MODEL_LABELLED,
user=Dataset.USER_MODEL_LABELLER,
is_labelled=False,
save=True,
)
num_scored += 1
return num_scored
def evaluate(self, x_seq, y_seq):
with self.graph.as_default():
x_eval, lengths = self.lang.tokenized_to_sequence(x_seq)
y_eval = utils.one_hot_encode_sequence(y_seq, self.classes)
return self.model.evaluate(x_eval, y_eval, verbose=0)
def evaluate(self, x_seq, y_seq):
with self.graph.as_default():
x_eval, lengths = self.lang.texts_to_sequence(x_seq)
y_eval = utils.one_hot_encode_sequence(
y_seq, valid_tokens=self.valid_outputs)
return self.model.evaluate(x_eval, y_eval)