def load_saved_model():
ner_model = NERCRF()
ner_model.build(model_info['sentence_len'],
model_info['word_len'],
model_info['num_of_labels'],
model_info['word_vocab'],
model_info['vocab_size'],
model_info['char_vocab_size'],
word_embedding_dims=model_info['word_embedding_dims'],
char_embedding_dims=model_info['char_embedding_dims'],
word_lstm_dims=model_info['word_lstm_dims'],
tagger_lstm_dims=model_info['tagger_lstm_dims'],
dropout=model_info['dropout'],
external_embedding_model=model_info['external_embedding_model'])
ner_model.load(args.model_path)
return ner_model
def load_saved_model():
ner_model = NERCRF()
ner_model.build(
model_info['sentence_len'],
model_info['word_len'],
model_info['num_of_labels'],
model_info['word_vocab'],
model_info['vocab_size'],
model_info['char_vocab_size'],
word_embedding_dims=model_info['word_embedding_dims'],
char_embedding_dims=model_info['char_embedding_dims'],
word_lstm_dims=model_info['word_lstm_dims'],
tagger_lstm_dims=model_info['tagger_lstm_dims'],
dropout=model_info['dropout'],
external_embedding_model=model_info['external_embedding_model'])
ner_model.load(args.model_path)
return ner_model
x_test, x_char_test, y_test = dataset.test_set
num_y_labels = len(dataset.y_labels) + 1
vocabulary_size = dataset.word_vocab_size
char_vocabulary_size = dataset.char_vocab_size
y_test = keras.utils.to_categorical(y_test, num_y_labels)
y_train = keras.utils.to_categorical(y_train, num_y_labels)
ner_model = NERCRF(use_cudnn=args.use_cudnn)
# pylint: disable=unexpected-keyword-arg
ner_model.build(
args.word_length,
num_y_labels,
vocabulary_size,
char_vocabulary_size,
word_embedding_dims=args.word_embedding_dims,
char_embedding_dims=args.character_embedding_dims,
tagger_lstm_dims=args.entity_tagger_lstm_dims,
dropout=args.dropout,
)
# initialize word embedding if external model selected
if args.embedding_model is not None:
embedding_model, _ = load_word_embeddings(args.embedding_model)
embedding_mat = get_embedding_matrix(embedding_model,
dataset.word_vocab)
ner_model.load_embedding_weights(embedding_mat)
train_inputs = [x_train, x_char_train]
test_inputs = [x_test, x_char_test]
x_test, x_char_test, y_test = dataset.test
num_y_labels = len(dataset.y_labels) + 1
vocabulary_size = dataset.word_vocab_size + 1
char_vocabulary_size = dataset.char_vocab_size + 1
y_test = to_categorical(y_test, num_y_labels)
y_train = to_categorical(y_train, num_y_labels)
ner_model = NERCRF()
ner_model.build(args.sentence_length,
args.word_length,
num_y_labels,
dataset.word_vocab,
vocabulary_size,
char_vocabulary_size,
word_embedding_dims=args.word_embedding_dims,
char_embedding_dims=args.character_embedding_dims,
word_lstm_dims=args.char_features_lstm_dims,
tagger_lstm_dims=args.entity_tagger_lstm_dims,
dropout=args.dropout,
external_embedding_model=args.embedding_model)
conll_cb = ConllCallback([x_test, x_char_test], y_test, dataset.y_labels,
batch_size=args.b)
ner_model.fit(x=[x_train, x_char_train], y=y_train,
batch_size=args.b,
epochs=args.e,
callbacks=[conll_cb],
validation=([x_test, x_char_test], y_test))
class NerApi(AbstractApi):
"""
Ner model API
"""
dir = path.dirname(path.realpath(__file__))
pretrained_model = path.join(dir, 'ner-pretrained', 'model.h5')
pretrained_model_info = path.join(dir, 'ner-pretrained', 'model_info.dat')
def __init__(self, ner_model=None, prompt=True):
self.model = None
self.model_info = None
self.model_path = NerApi.pretrained_model
self.model_info_path = NerApi.pretrained_model_info
self._download_pretrained_model(prompt)
def encode_word(self, word):
return self.model_info['word_vocab'].get(word, 1.0)
def encode_word_chars(self, word):
return [self.model_info['char_vocab'].get(c, 1.0) for c in word]
def encode_input(self, text_arr):
sentence = []
sentence_chars = []
for word in text_arr:
sentence.append(self.encode_word(word))
sentence_chars.append(self.encode_word_chars(word))
encoded_sentence = pad_sequences(
[np.asarray(sentence)], maxlen=self.model_info['sentence_len'])
chars_padded = pad_sequences(sentence_chars,
maxlen=self.model_info['word_len'])
if self.model_info['sentence_len'] - chars_padded.shape[0] > 0:
chars_padded = np.concatenate((np.zeros(
(self.model_info['sentence_len'] - chars_padded.shape[0],
self.model_info['word_len'])), chars_padded))
encoded_chars = chars_padded.reshape(1,
self.model_info['sentence_len'],
self.model_info['word_len'])
return encoded_sentence, encoded_chars
def _prompt(self):
response = input(
'\nTo download \'{}\', please enter YES: '.format('ner'))
res = response.lower().strip()
if res == "yes" or (len(res) == 1 and res == 'y'):
print('Downloading {}...'.format('ner'))
responded_yes = True
else:
print('Download declined. Response received {} != YES|Y. '.format(
res))
responded_yes = False
return responded_yes
def _download_pretrained_model(self, prompt=True):
"""Downloads the pre-trained BIST model if non-existent."""
dir_path = path.join(self.dir, 'ner-pretrained')
if not path.isfile(path.join(dir_path, 'model.h5')):
print(
'The pre-trained models to be downloaded for the NER dataset '
'are licensed under Apache 2.0. By downloading, you accept the terms '
'and conditions provided by the license')
makedirs(dir_path, exist_ok=True)
if prompt is True:
agreed = self._prompt()
if agreed is False:
sys.exit(0)
download_unlicensed_file(
'http://nervana-modelzoo.s3.amazonaws.com/NLP/NER/',
'model.h5', self.model_path)
download_unlicensed_file(
'http://nervana-modelzoo.s3.amazonaws.com/NLP/NER/',
'model_info.dat', self.model_info_path)
print('Done.')
def load_model(self):
with open(self.model_info_path, 'rb') as fp:
self.model_info = pickle.load(fp)
self.model = NERCRF()
self.model.build(
self.model_info['sentence_len'],
self.model_info['word_len'],
self.model_info['num_of_labels'],
self.model_info['word_vocab'],
self.model_info['vocab_size'],
self.model_info['char_vocab_size'],
word_embedding_dims=self.model_info['word_embedding_dims'],
char_embedding_dims=self.model_info['char_embedding_dims'],
word_lstm_dims=self.model_info['word_lstm_dims'],
tagger_lstm_dims=self.model_info['tagger_lstm_dims'],
dropout=self.model_info['dropout'],
external_embedding_model=self.
model_info['external_embedding_model'])
self.model.load(self.model_path)
def pretty_print(self, text, tags):
tags_str = [
self.model_info['labels_id_to_word'].get(t, None) for t in tags[0]
][-len(text):]
mapped = [{
'index': idx,
'word': el,
'label': tags_str[idx]
} for idx, el in enumerate(text)]
counter = 0
ents = []
spans = []
for obj in mapped:
if (obj['label'] != 'O'):
spans.append({
'start': counter,
'end': (counter + len(obj['word'])),
'type': obj['label']
})
counter += len(obj['word']) + 1
ents = dict((obj['type'].lower(), obj) for obj in spans).keys()
ret = {}
ret['doc_text'] = ' '.join(text)
ret['annotation_set'] = list(ents)
ret['spans'] = spans
ret['title'] = 'None'
return {"doc": ret, 'type': 'high_level'}
def process_text(self, text):
input_text = ' '.join(text.strip().split())
return nlp.tokenize(input_text)
def inference(self, doc):
text_arr = self.process_text(doc)
words, chars = self.encode_input(text_arr)
tags = self.model.predict([words, chars])
tags = tags.argmax(2)
return self.pretty_print(text_arr, tags)
