class Kashgari:
def __init__(self):
self.model = None
self.chunk_size = 100
self.set_features_numeric = dict()
self.set_features_text = dict()
def prepare_data_fit(self, tokens, tags, chunk_size, overlap=10):
text_list = []
first_of_p_list = []
tag_list = []
buffer_text = []
buffer_first_of_p = []
buffer_tag = []
text_features = set("token")
numeric_features = set("first_of_p")
self.set_features_numeric = dict()
for doc, doc_tags in zip(tokens, tags):
for token, tag in zip(doc, doc_tags):
features = agregado(token, simple_features=True)
buffer_text.append(features['token'])
buffer_first_of_p.append(
'2' if features['first_of_p'] else '1')
buffer_tag.append(tag)
if len(buffer_text) > chunk_size:
text_list.append(buffer_text)
first_of_p_list.append(buffer_first_of_p)
tag_list.append(buffer_tag)
# Zerar
buffer_text = []
buffer_first_of_p = []
buffer_tag = []
print("Processed doc")
if len(buffer_text) >= 0:
text_list.append(buffer_text)
first_of_p_list.append(buffer_first_of_p)
tag_list.append(buffer_tag)
results = (text_list, first_of_p_list)
return results, tag_list
def prepare_data_predict(self, tokens, chunk_size):
text_list = []
first_of_p_list = []
buffer_text = []
buffer_first_of_p = []
for token in tokens:
features = agregado(token, simple_features=True)
buffer_text.append(features['token'])
buffer_first_of_p.append('2' if features['first_of_p'] else '1')
if len(buffer_text) >= chunk_size:
text_list.append(buffer_text)
first_of_p_list.append(buffer_first_of_p)
# Zerar
buffer_text = []
buffer_first_of_p = []
if len(buffer_text) > 0:
text_list.append(buffer_text)
first_of_p_list.append(buffer_first_of_p)
results = (text_list, first_of_p_list)
return results
def train(self, tokens, tags):
x, y = self.prepare_data_fit(tokens, tags, chunk_size=self.chunk_size)
text_embedding = BareEmbedding(task=kashgari.LABELING,
sequence_length=self.chunk_size)
first_of_p_embedding = NumericFeaturesEmbedding(
feature_count=2,
feature_name='first_of_p',
sequence_length=self.chunk_size)
stack_embedding = StackedEmbedding(
[text_embedding, first_of_p_embedding])
stack_embedding.analyze_corpus(x, y)
from kashgari.tasks.labeling import BiLSTM_Model, BiLSTM_CRF_Model
self.model = BiLSTM_CRF_Model(embedding=stack_embedding)
self.model.fit(x, y, batch_size=1, epochs=20)
def predict(self, tokens):
import itertools
results = []
for doc in tokens:
x = self.prepare_data_predict(doc, chunk_size=self.chunk_size)
predicted = self.model.predict(x)
x_list = list(itertools.chain.from_iterable(x[0]))
predicted_unified = list(itertools.chain.from_iterable(predicted))
predicted_truncated = predicted_unified[:len(doc)]
print(
f"len doc{len(doc)} | x_list{len(x_list)} |len predicted_unified{len(predicted_unified)} |len predicted_truncated{len(predicted_truncated)} |"
)
results.append(predicted_unified[:len(doc)])
return results
text = [[0.9, 0.1, 0.1], [0.9, 0.1, 0.1], [0.1, 0.8, 0.1], [0.1, 0.8, 0.1],
[0.1, 0.8, 0.1]]
label = [
'B-Category', 'I-Category', 'B-ProjectName', 'I-ProjectName',
'I-ProjectName'
]
text_list = [text] * 100
label_list = [label] * 100
SEQUENCE_LEN = 80
# You can use WordEmbedding or BERTEmbedding for your text embedding
bare_embedding = DirectEmbedding(task=kashgari.RAW_LABELING,
sequence_length=SEQUENCE_LEN,
embedding_size=3)
#bare_embedding = BareEmbedding(task=kashgari.LABELING, sequence_length=SEQUENCE_LEN)
x = (text_list)
y = label_list
bare_embedding.analyze_corpus(x, y)
# Now we can embed with this stacked embedding layer
# We can build any labeling model with this embedding
from kashgari.tasks.labeling import BiLSTM_Model, BiLSTM_CRF_Model
model = BiLSTM_CRF_Model(embedding=bare_embedding)
model.fit(x, y, batch_size=1, epochs=3)
print(model.predict(x))
#print(model.predict_entities(x))
class BertProsody:
""" 目前只支持长度50，输入字符数49 + 终结符 """
def __init__(self):
self.model, self.model_dir, self.model_path = None, None, None
self.sess = None
return
def initial_model(self, bert_model_path, psd_model_path):
print('=============init bert model=========================')
print("bert model path:", bert_model_path)
print("crf model path:", psd_model_path)
self.sess = tf.Session()
set_session(self.sess)
self.model_dir = os.path.dirname(os.path.dirname(psd_model_path))
self.model_path = psd_model_path
data_path = os.path.join(self.model_dir, "feature_psd.pkl")
train_data, train_label, test_data, test_label = \
pickle.load(open(data_path, 'rb'))
bert_embed = BERTEmbedding(bert_model_path, task=kashgari.LABELING,
sequence_length=50)
self.model = BiLSTM_CRF_Model(bert_embed)
self.model.build_model(x_train=train_data, y_train=train_label,
x_validate=test_data, y_validate=test_label)
self.model.compile_model()
self.model.tf_model.load_weights(psd_model_path)
print('=============bert model loaded=========================')
return
def _write_dict(self):
label_path = os.path.join(self.model_dir, "idx2label.txt")
with open(label_path, "w", encoding="utf-8") as fr:
for key, value in self.model.embedding.label2idx.items():
fr.write("{} {}\n".format(value, key))
token_path = os.path.join(self.model_dir, "token2idx.txt")
with open(token_path, "w", encoding="utf-8") as fr:
for key, value in self.model.embedding.token2idx.items():
if len(key) > 0:
fr.write("{} {}\n".format(key, value))
def predict(self, sentence_list):
""" 通过句子预测韵律，标点断开 """
bert_input = []
for sent in sentence_list:
assert len(sent) < 50
bert_input.append([c for c in sent])
print("bert-input:", bert_input)
prosody = self.model.predict(bert_input)
return prosody
def compute_embed(self, sentence_list):
bert_input = [[c for c in sent] for sent in sentence_list]
print("bert-input:", bert_input)
tensor = self.model.embedding.process_x_dataset(bert_input)
res = self.model.tf_model.predict(tensor)
import numpy as np
print("debug:", np.shape(res), res[0])
return tensor
def save_pb(self):
self._write_dict()
pb_dir = os.path.join(self.model_dir, "pb")
os.makedirs(pb_dir, exist_ok=True)
# [print(n.name) for n in tf.get_default_graph().as_graph_def().node]
h5_to_pb(self.model.tf_model, pb_dir, self.sess, "model_psd.pb",
["output_psd"])
return
@staticmethod
def change_by_rules(old_pairs):
""" 强制规则:
1. 逗号之前是#3，句号之前是#4
2. 其他位置，#3 -> #2
"""
new_pairs = []
for i, (char, ph, psd) in enumerate(old_pairs[0:-1]):
next_char, _, _ = old_pairs[i+1]
if next_char == "，":
new_pairs.append((char, ph, "3"))
elif next_char in ["。", "?", "!"]:
new_pairs.append((char, ph, "4"))
else:
if psd == "3":
new_pairs.append((char, ph, "2"))
else:
new_pairs.append((char, ph, psd))
new_pairs.append(old_pairs[-1])
return new_pairs
class BertPolyPhone:
""" 拼音预测主类"""
def __init__(self):
super().__init__()
self.poly_dict = dict()
poly_dict_path = "/data1/liufeng/synthesis/frontend/data/simple_poly_dict"
for line in read_lines(poly_dict_path):
line = line.replace(" ", "").replace("*", "")
key = line.split(":")[0]
value = line.split(":")[1].split(",")
self.poly_dict[key] = value
self.model, self.model_dir = None, None
self.sess = None
def inialize_model(self, bert_model_path, poly_model_path):
print('=============init phone model=========================')
print("bert model path:", bert_model_path)
print("crf model path:", poly_model_path)
# 需要训练数据的路径构建字典
self.sess = tf.Session()
set_session(self.sess)
self.model_dir = os.path.dirname(os.path.dirname(poly_model_path))
data_path = os.path.join(self.model_dir, "feature.pkl")
train_data, train_label, test_data, test_label = \
pickle.load(open(data_path, 'rb'))
bert_embed = BERTEmbedding(bert_model_path, task=kashgari.LABELING,
sequence_length=50)
self.model = BiLSTM_CRF_Model(bert_embed)
self.model.build_model(x_train=train_data, y_train=train_label,
x_validate=test_data, y_validate=test_label)
self.model.compile_model()
self.model.tf_model.load_weights(poly_model_path)
print('=============successful loaded=========================')
def _lookup_dict(self, bert_result, pred_ph_pairs):
"""查字典的方法对拼音进行修正 """
# todo: 如果词在词典中，不用bert的结果。
bert_phone_result = []
for index_c, (char, ph, _) in enumerate(pred_ph_pairs):
if char in self.poly_dict.keys():
# 如果bert预测结果不在多音字字典中，就是预测结果跑偏了
if bert_result[index_c] not in self.poly_dict[char]:
bert_phone_result.append((char, ph))
else:
bert_result[index_c] = split_phone_format(bert_result[index_c])
bert_phone_result.append((char, bert_result[index_c]))
if ph != bert_result[index_c]:
print("using bert result {}:{} instead of {}".format(
char, bert_result[index_c], ph))
else:
bert_phone_result.append((char, ph))
return bert_phone_result
def predict(self, sentence_list):
""" 通过句子预测韵律，标点断开 """
bert_input = []
for sent in sentence_list:
assert len(sent) < 50
bert_input.append([c for c in sent])
print("bert-input:", bert_input)
prosody = self.model.predict(bert_input)
return prosody
def save_pb(self):
self._write_dict()
pb_dir = os.path.join(self.model_dir, "pb")
os.makedirs(pb_dir, exist_ok=True)
h5_to_pb(self.model.tf_model, pb_dir, self.sess, "model_phone.pb",
["output_phone"])
return
def _write_dict(self):
label_path = os.path.join(self.model_dir, "pb/phone_idx2label.txt")
with open(label_path, "w", encoding="utf-8") as fr:
for key, value in self.model.embedding.label2idx.items():
fr.write("{} {}\n".format(value, key))
print("write {}".format(label_path))
token_path = os.path.join(self.model_dir, "pb/phone_token2idx.txt")
with open(token_path, "w", encoding="utf-8") as fr:
for key, value in self.model.embedding.token2idx.items():
if len(key) > 0:
fr.write("{} {}\n".format(key, value))
print("write {}".format(token_path))
return
def compute_embed(self, sentence_list):
bert_input = [[c for c in sent] for sent in sentence_list]
print("bert-input:", bert_input)
import numpy as np
tensor = self.model.embedding.process_x_dataset(bert_input)
print("debug:", np.shape(tensor), tensor)
res = self.model.tf_model.predict(tensor)
import numpy as np
print("debug:", np.shape(res), res[0][0: len(sentence_list[0]+1)])
return tensor
@staticmethod
def _merge_eng_char(bert_phone_result, dict_phone_pairs):
from src.utils import check_all_chinese
index = 0
new_bert_phone = []
for word, _, _ in dict_phone_pairs:
if (not check_all_chinese(word)) and len(word) > 1:
new_bert_phone.append(bert_phone_result[index])
index += len(word)
else:
new_bert_phone.append(bert_phone_result[index])
index += 1
return new_bert_phone
def modify_result(self, bert_result, dict_phone_pairs):
bert_result = self._merge_eng_char(bert_result, dict_phone_pairs)
bert_phone_pairs = self._lookup_dict(bert_result, dict_phone_pairs)
phone_pairs = bert_phone_pairs
# phone_pairs = change_yi(phone_pairs)
# phone_pairs = change_bu(phone_pairs)
phone_pairs = sandhi(phone_pairs)
bert_result = [ph for _, ph in phone_pairs]
chars = "".join([c for c, _ in phone_pairs])
bert_result = change_qingyin(bert_result, chars)
return bert_result
# ou o GloVE-300 do http://nilc.icmc.usp.br/embeddings se não der certo
# 2 - Ver como fazer o Predict. Temos que processar a frase para ficar igual a deles.
# Eles usam um PunktSentenceTokenizer com um abbrev_list. Esses scripts estao na pasta leNer-dataset.
# 3 - Ver como integrar esse codigo com o webstruct atual
# 4 - Seria uma boa ideia ter uma interface tipo o Broka. Para que existesse a lista de arquivos, e que
# pudesse abrir para re-treinar, abrindo com o plugin de Ramon.
# Uma ideia seria ate converter o dataset deles atual para o formato do broka hoje em Html ( pode ser algo simples, como colocar cada paragrafo como um p)
# 5 - Fazer a persistencia ( O kashgari tem um metodo save/load)
# 2 - Aumentar epochs para treinar
# You can use WordEmbedding or BERTEmbedding for your text embedding
text_embedding = BareEmbedding(task=kashgari.LABELING)
text_embedding.analyze_corpus(tokens, labels)
# Now we can embed with this stacked embedding layer
# We can build any labeling model with this embedding
from kashgari.tasks.labeling import BiLSTM_CRF_Model
model = BiLSTM_CRF_Model(embedding=text_embedding)
model.fit(tokens, labels, batch_size=8, epochs=10)
print(model.predict(tokens))
# print(model.predict_entities(x))