def albert_process_data(self, mode='part'):
_token_dict = load_vocab(self.albert_dict_path) # 读取字典
# 只取涉及数据集中出现的字
if mode == 'part':
train_df = pd.read_csv(self.train_data_path,
names=['seq1', 'seq2', 'label'])
valid_df = pd.read_csv(self.dev_data_path,
names=['seq1', 'seq2', 'label'])
test_df = pd.read_csv(self.test_data_path,
names=['seq1', 'seq2', 'label'])
# total data
tmp_df = pd.concat([train_df, valid_df, test_df])
chars = defaultdict(int)
for _, tmp_row in tmp_df.iterrows():
for tmp_char in tmp_row.seq1:
chars[tmp_char] += 1
for tmp_char in tmp_row.seq2:
chars[tmp_char] += 1
# 过滤低频字
chars = {i: j for i, j in chars.items() if j >= 4}
self.token_dict, self.keep_words = {}, [] # keep_words是在bert中保留的字表
# 保留特殊字符
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
# 字典只保留数据中出现的高频字
for c in chars:
if c in _token_dict:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
elif mode == 'full':
self.token_dict, self.keep_words = _token_dict, []
for k in self.token_dict:
self.keep_words.append(self.token_dict[k])
self.tokenizer = SimpleTokenizer(self.token_dict) # 建立分词器
def process_data(train_file, dev_file, test_file):
chars = set()
train_datas = read_data(train_file)
dev_datas = read_data(dev_file)
test_datas = read_data(test_file)
for text1, text2, label in train_datas + dev_datas:
chars.update(set(text1))
chars.update(set(text2))
_token_dict = load_vocab(dict_path) # 读取词典
token_dict, keep_words = {}, []
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
token_dict[c] = len(token_dict)
keep_words.append(_token_dict[c])
for c in chars:
if c in _token_dict:
token_dict[c] = len(token_dict)
keep_words.append(_token_dict[c])
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
with open(os.path.join(model_save_path, 'tokenizer.pkl'), "wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'), "wb") as f:
pickle.dump(keep_words, f)
return train_datas, dev_datas, test_datas, tokenizer, keep_words
def save_vocab(self, input_data, incremental_train=False):
relationships = set()
chars = set()
for (text, triple), (entity_lists, rel) in input_data:
chars.update(set(text))
relationships.add(rel)
relationships.update(set(p for s, p, o in triple))
token_dict = load_vocab(dict_path) # 读取词典
keep_words = list(set(token_dict.values()))
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
# keep_flags = ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']
rel2id = {rel: _id + 1 for _id, rel in enumerate(sorted(relationships))}
rel2id['unk'] = 0
if not incremental_train:
with open(os.path.join(model_save_path, 'tokenizer.pkl'), "wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'), "wb") as f:
pickle.dump(keep_words, f)
with open(os.path.join(model_save_path, 'rel2id.pkl'), "wb") as f:
pickle.dump(rel2id, f)
self.tokenizer, self.keep_words, self.rel2id = tokenizer, keep_words, rel2id
return tokenizer, keep_words, rel2id
def __init__(self,
initial_model=True,
model_path=os.path.join(CONFIG['model_dir'], 'albert.h5')):
self.initial_model = initial_model
token_dict = load_vocab(DICT_PATH)
self.tokenizer = SimpleTokenizer(token_dict)
self.model_path = model_path
if initial_model:
self.albert_model = load_pretrained_model(
CONFIG_PATH,
CHECKPOINT_PATH,
# keep_words=keep_words,
albert=True)
else:
self.load(model_path)
for l in self.albert_model.layers:
l.trainable = True
def __init__(self, batch_size=32, train=False):
self.batch_size = batch_size
if train:
chars = set()
train_datas = read_datas(TRAIN_DATA_FILE)
dev_datas = read_datas(DEV_DATA_FILE)
test_datas = read_datas(TEST_DATA_FILE)
for text1, text2, label in itertools.chain(train_datas, dev_datas):
chars.update(set(text1))
chars.update(set(text2))
_token_dict = load_vocab(dict_path) # 读取词典
self.token_dict, self.keep_words = {}, []
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
for c in chars:
if c in _token_dict:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
self.tokenizer = SimpleTokenizer(self.token_dict) # 建立分词器
with open(os.path.join(model_save_path, 'tokenizer.pkl'),
"wb") as f:
pickle.dump(self.tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'),
"wb") as f:
pickle.dump(self.keep_words, f)
else:
with open(os.path.join(model_save_path, 'tokenizer.pkl'),
"rb") as f:
self.tokenizer = pickle.load(f)
with open(os.path.join(model_save_path, 'keep_words.pkl'),
"rb") as f:
self.keep_words = pickle.load(f)
self.model = self.make_model()
def get_correct_fn():
save_path = join(MODEL_PATH, 'detect')
token_dict = joblib.load(
join(MODEL_PATH, 'train_pre_for_error_detect', 'token_dict.joblib'))
tokenizer = SimpleTokenizer(token_dict)
keep_words = joblib.load(
join(MODEL_PATH, 'train_pre_for_error_detect', 'keep_words.joblib'))
model = DetectModel(keep_words=keep_words)
model.compile()
model.model.load_weights(join(save_path, 'weights.hdf5'))
checker = Statistics()
def correct(error_text):
text_tokens = tokenizer.tokenize(error_text, False,
False)[:ec_cfg.max_seq_len - 2]
tokens = list()
tokens.append("[CLS]")
for token in text_tokens:
tokens.append(token)
tokens.append("[SEP]")
input_ids = [
token_dict[c] if c in token_dict.keys() else token_dict['[UNK]']
for c in tokens
]
while len(input_ids) < ec_cfg.max_seq_len:
input_ids.append(0)
seg_ids = np.zeros_like(input_ids, dtype=np.int)
ids, segs = [input_ids], [seg_ids]
res = model.model.predict([ids, segs])[0][1:-1]
begins_pred = []
lengths_pred = []
this_len = 0
for i, r in enumerate(res):
if np.argmax(r) > 0:
if this_len == 0:
begins_pred.append(i)
this_len += 1
else:
if this_len > 0:
lengths_pred.append(this_len)
this_len = 0
else:
if this_len > 0:
lengths_pred.append(this_len)
res_str = checker.correct(error_text, begins_pred, lengths_pred)
return res_str
return correct
def process_data(neg_file='datasets/neg.xls', pos_file='datasets/pos.xls'):
neg = pd.read_excel(neg_file, header=None)
pos = pd.read_excel(pos_file, header=None)
chars = {}
data = []
for d in neg[0]:
data.append((d, 0))
for c in d:
chars[c] = chars.get(c, 0) + 1
for d in pos[0]:
data.append((d, 1))
for c in d:
chars[c] = chars.get(c, 0) + 1
chars = {i: j for i, j in chars.items() if j >= 4}
_token_dict = load_vocab(dict_path) # 读取词典
token_dict, keep_words = {}, set()
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
token_dict[c] = len(token_dict)
keep_words.add(_token_dict[c])
for c in chars:
if c in _token_dict:
token_dict[c] = len(token_dict)
keep_words.add(_token_dict[c])
keep_words.add(max(keep_words) + 1)
keep_words = list(keep_words)
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
with open(os.path.join(model_save_path, 'tokenizer.pkl'), "wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'), "wb") as f:
pickle.dump(keep_words, f)
if not os.path.exists('./random_order.json'):
random_order = [i for i in range(len(data))]
random.shuffle(random_order)
json.dump(random_order, open('./random_order.json', 'w'), indent=4)
else:
random_order = json.load(open('./random_order.json'))
# 按照9:1的比例划分训练集和验证集
train_data = [data[j] for i, j in enumerate(random_order) if i % 10 != 0]
valid_data = [data[j] for i, j in enumerate(random_order) if i % 10 == 0]
return train_data, valid_data, tokenizer, keep_words
def save_vocab(self, save_path, process_data):
chars = set()
relationships = set()
for text, relationship in process_data:
words = split_text(text)
chars.update(set(words))
relationships.add(relationship)
token_dict = load_vocab(dict_path) # 读取词典
keep_chars = ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']
for char in chars:
if not token_dict.get(char):
# token_dict[char] = len(token_dict)
keep_chars.append(char)
# for char in keep_chars:
# if not token_dict.get(char):
# token_dict[char] = len(token_dict)
keep_words = list(set(token_dict.values()))
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
word2id = {
word: id_ + len(keep_chars)
for id_, word in enumerate(chars)
}
for _id, word in enumerate(keep_chars):
word2id[word] = _id
rel2id = {rel: _id for _id, rel in enumerate(relationships)}
with open(os.path.join(model_save_path, 'tokenizer.pkl'), "wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'), "wb") as f:
pickle.dump(keep_words, f)
with open(os.path.join(save_path, 'word2id.pkl'), "wb") as f:
pickle.dump(word2id, f)
with open(os.path.join(save_path, 'rel2id.pkl'), "wb") as f:
pickle.dump(rel2id, f)
self.tokenizer, self.word2id, self.rel2id = tokenizer, word2id, rel2id
return tokenizer, keep_words, word2id, rel2id
def save_vocab(self, save_path, process_data):
flags = set()
relationships = set()
for old_word_flag, relationship in process_data:
word_flag = []
for word, flag in old_word_flag:
# if flag[0] == 'B':
# flag = 'B-Shiyi'
# elif flag[0] == 'I':
# flag = 'I-Shiyi'
word_flag.append([word, flag])
flags.update(set(flag for word, flag in word_flag))
relationships.add(relationship)
token_dict = load_vocab(dict_path) # 读取词典
keep_words = list(set(token_dict.values()))
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
keep_flags = ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']
flag2id = {
label: id_ + len(keep_flags)
for id_, label in enumerate(
sorted(flags, key=lambda x: 0 if x == 'O' else 1))
}
for flag_id, flag in enumerate(keep_flags):
flag2id[flag] = flag_id
rel2id = {rel: _id for _id, rel in enumerate(relationships)}
with open(os.path.join(model_save_path, 'tokenizer.pkl'), "wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'), "wb") as f:
pickle.dump(keep_words, f)
with open(os.path.join(save_path, 'flag2id.pkl'), "wb") as f:
pickle.dump(flag2id, f)
with open(os.path.join(save_path, 'rel2id.pkl'), "wb") as f:
pickle.dump(rel2id, f)
self.tokenizer, self.flag2id, self.rel2id = tokenizer, flag2id, rel2id
return tokenizer, keep_words, flag2id, rel2id
def save_vocab(self, model_save_path, process_data):
chars = set()
labels = set()
for char_labels in process_data:
for char, label in char_labels:
chars.add(char)
labels.add(label)
_token_dict = load_vocab(dict_path) # 读取词典
token_dict, keep_words = {}, set()
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
token_dict[c] = len(token_dict)
keep_words.add(_token_dict[c])
for c in chars:
if c in _token_dict:
token_dict[c] = len(token_dict)
keep_words.add(_token_dict[c])
keep_words.add(max(keep_words) + 1)
keep_words = list(keep_words)
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
with open(os.path.join(model_save_path, 'tokenizer.pkl'), "wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'), "wb") as f:
pickle.dump(keep_words, f)
# print('labels={}'.format(labels))
# sorted: 保证 非实体词 O 的id为0
self.label2id = {
label: id_
for id_, label in enumerate(
sorted(labels, key=lambda x: 0 if x == 'O' else 1))
}
print('label2id: {}'.format(self.label2id))
with open(os.path.join(model_save_path, 'label2id.pkl'), "wb") as f:
pickle.dump(self.label2id, f)
self.keep_words = keep_words
self.tokenizer = tokenizer
def save_word2id_etc(self, datas, incremental_train=False):
label_set = set()
_token_dict = load_vocab(dict_path) # 读取词典
# token_dict, keep_words = {}, set()
token_dict = copy.deepcopy(_token_dict)
# for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
# token_dict[c] = len(token_dict)
# keep_words.add(_token_dict[c])
for chars, label in datas:
label_set.add(label)
# for c in chars:
# if c in _token_dict:
# token_dict[c] = len(token_dict)
# keep_words.add(_token_dict[c])
# keep_words.add(max(keep_words)+1)
# keep_words = list(keep_words)
keep_words = list(set(token_dict.values()))
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
label2id = {lab: i for i, lab in enumerate(list(label_set))}
if not incremental_train:
with open(os.path.join(model_save_path, 'tokenizer.pkl'),
"wb") as f:
pickle.dump(tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'),
"wb") as f:
pickle.dump(keep_words, f)
with open(os.path.join(model_save_path, 'label2id.pkl'),
"wb") as f:
pickle.dump(label2id, f)
return tokenizer, keep_words, label2id
import tensorflow as tf
from bert4keras.bert import load_pretrained_model
from bert4keras.utils import SimpleTokenizer, load_vocab
import numpy as np
gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
print("Name:", gpu.name, " Type:", gpu.device_type)
tf.config.experimental.set_virtual_device_configuration(
gpus[0],
[tf.config.experimental.VirtualDeviceConfiguration(memory_limit=1024)])
print(tf.__version__)
base_path = 'D:\AI\Data\chinese_L-12_H-768_A-12\\'
config_path = base_path + 'bert_config.json'
checkpoint_path = base_path + 'bert_model.ckpt'
dict_path = base_path + 'vocab.txt'
token_dict = load_vocab(dict_path) # 读取词典
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
model = load_pretrained_model(config_path, checkpoint_path) # 建立模型，加载权重
# 编码测试
token_ids, segment_ids = tokenizer.encode(u'语言模型')
print(model.predict([np.array([token_ids]), np.array([segment_ids])]))
import tensorflow as tf
from bert4keras.bert import load_pretrained_model
from bert4keras.utils import SimpleTokenizer, load_vocab
import numpy as np
gpus = tf.config.experimental.list_physical_devices('GPU')
tf.config.experimental.set_virtual_device_configuration(
gpus[0],
[tf.config.experimental.VirtualDeviceConfiguration(memory_limit=2048)])
base_path = 'D:\AI\Data\\albert_large_zh\\'
config_path = base_path + 'albert_config_large.json'
checkpoint_path = base_path + 'albert_model.ckpt'
dict_path = base_path + 'vocab.txt'
token_dict = load_vocab(dict_path) # 读取词典
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
model = load_pretrained_model(config_path, checkpoint_path,
albert=True) # 建立模型，加载权重
token_ids, segment_ids = tokenizer.encode(u'科学技术是第一生产力')
# mask掉“技术”
token_ids[3] = token_ids[4] = token_dict['[MASK]']
# 用mlm模型预测被mask掉的部分
probas = model.predict([np.array([token_ids]), np.array([segment_ids])])[0]
print(tokenizer.decode(probas[3:5].argmax(axis=1))) # 结果正是“技术”权
class Albert(object):
def __init__(self,
mode='inference',
mode_='part',
model_name=None,
dataset_name=None):
self.maxlen = 32
self.albert_config_path = '/Data/public/Bert/albert_tiny_489k/albert_config_tiny.json'
self.albert_checkpoint_path = '/Data/public/Bert/albert_tiny_489k/albert_model.ckpt'
self.albert_dict_path = '/Data/public/Bert/albert_tiny_489k/vocab.txt'
self.train_data_path = 'data/train_{}.csv'.format(dataset_name)
self.dev_data_path = 'data/dev_{}.csv'.format(dataset_name)
self.test_data_path = 'data/test_{}.csv'.format(dataset_name)
# albert_tiny_250k.h5 挺好的
# self.restore_model_path = 'saved_models/test_albert_tiny_{}.h5'.format(model_name)
self.restore_model_path = '/Data/models/{}'.format(model_name)
# albert
self.albert_process_data(mode_)
if mode == 'train':
self.model = self._get_model()
self.train()
elif mode == 'inference':
self._init_model()
# todo keep words 工业场景下需要remove
def albert_process_data(self, mode='part'):
_token_dict = load_vocab(self.albert_dict_path) # 读取字典
# 只取涉及数据集中出现的字
if mode == 'part':
train_df = pd.read_csv(self.train_data_path,
names=['seq1', 'seq2', 'label'])
valid_df = pd.read_csv(self.dev_data_path,
names=['seq1', 'seq2', 'label'])
test_df = pd.read_csv(self.test_data_path,
names=['seq1', 'seq2', 'label'])
# total data
tmp_df = pd.concat([train_df, valid_df, test_df])
chars = defaultdict(int)
for _, tmp_row in tmp_df.iterrows():
for tmp_char in tmp_row.seq1:
chars[tmp_char] += 1
for tmp_char in tmp_row.seq2:
chars[tmp_char] += 1
# 过滤低频字
chars = {i: j for i, j in chars.items() if j >= 4}
self.token_dict, self.keep_words = {}, [] # keep_words是在bert中保留的字表
# 保留特殊字符
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
# 字典只保留数据中出现的高频字
for c in chars:
if c in _token_dict:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
elif mode == 'full':
self.token_dict, self.keep_words = _token_dict, []
for k in self.token_dict:
self.keep_words.append(self.token_dict[k])
self.tokenizer = SimpleTokenizer(self.token_dict) # 建立分词器
# data pre-processing operation
def _data_preprocessing(self, sentence1, sentence2):
X1, X2 = [], []
for tmp_sent1, tmp_sent2 in zip(sentence1, sentence2):
x1, x2 = self.tokenizer.encode(first=tmp_sent1[:self.maxlen],
second=tmp_sent2[:self.maxlen])
X1.append(x1)
X2.append(x2)
X1 = self._seq_padding(X1)
X2 = self._seq_padding(X2)
# X1 = pad_sequences(X1, maxlen=67, padding='post', truncating='post')
# X2 = pad_sequences(X2, maxlen=67, padding='post', truncating='post')
return X1, X2
def _seq_padding(self, X, padding=0):
L = [len(x) for x in X]
ML = max(L)
padded_sent = np.array([
np.concatenate([x, [padding] *
(ML - len(x))]) if len(x) < ML else x for x in X
])
return padded_sent
# prepare data for training
def _prepare_data(self, data_path):
data = pd.read_csv(data_path)
sent_1 = data['sentence1'].values
sent_2 = data['sentence2'].values
label = data['label'].values
X1_pad, X2_pad = self._data_preprocessing(sent_1, sent_2)
# X1 = np.vstack((X1_pad, X2_pad))
# X2 = np.vstack((X2_pad, X1_pad))
# y_train = np.hstack((label, label))
return X1_pad, X2_pad, label
# albert for Semantic matching, model architecture
def _get_model(self):
model = load_pretrained_model(
self.albert_config_path,
self.albert_checkpoint_path,
keep_words=self.keep_words, # 只保留keep_words中的字，精简原字表
albert=True)
output = Lambda(lambda x: x[:, 0])(model.output)
output = Dense(1, activation='sigmoid')(output)
model = Model(model.input, output)
return model
# model training operation
def train(self):
# train_data
train_x1, train_x2, train_label = self._prepare_data(
self.train_data_path)
# dev_data
dev_x1, dev_x2, dev_label = self._prepare_data(self.dev_data_path)
checkpoint = ModelCheckpoint(self.restore_model_path,
monitor='val_accuracy',
verbose=0,
save_best_only=True,
save_weights_only=False)
early_stop = EarlyStopping(monitor='val_accuracy',
patience=3,
verbose=0,
mode='auto',
baseline=None,
restore_best_weights=True)
self.model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-4), # 用足够小的学习率
metrics=['accuracy'])
self.model.summary()
self.model.fit(x=[train_x1, train_x2],
y=train_label,
batch_size=64,
epochs=10,
verbose=1,
callbacks=[checkpoint, early_stop],
validation_data=([dev_x1, dev_x2], dev_label))
# model predict operation
def predict(self, sentence1, sentence2):
X1, X2 = self._data_preprocessing(sentence1, sentence2)
y_pred = self.model.predict([X1, X2], batch_size=1024)
return y_pred
def test(self):
self.model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-4), # 用足够小的学习率
metrics=['accuracy'])
# test_data
test_x1, test_x2, test_label = self._prepare_data(self.dev_data_path)
test_loss, test_acc = self.model.evaluate(x=[test_x1, test_x2],
y=test_label)
print('test loss: {}'.format(test_loss))
print('test acc: {}'.format(test_acc))
def _init_model(self):
self.model = load_model(self.restore_model_path)
sentence1 = '干嘛呢'
sentence2 = '你是机器人'
print('model albert loaded succeed. ({})'.format(
self.predict([sentence1], [sentence2]).item()))
chars = {i: j for i, j in chars.items() if j >= 4}
_token_dict = load_vocab(dict_path) # 读取词典
token_dict, keep_words = {}, []
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
token_dict[c] = len(token_dict)
keep_words.append(_token_dict[c])
for c in chars:
if c in _token_dict:
token_dict[c] = len(token_dict)
keep_words.append(_token_dict[c])
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
if not os.path.exists('./random_order.json'):
random_order = range(len(data))
np.random.shuffle(random_order)
json.dump(random_order, open('./random_order.json', 'w'), indent=4)
else:
random_order = json.load(open('./random_order.json'))
# 按照9:1的比例划分训练集和验证集
train_data = [data[j] for i, j in enumerate(random_order) if i % 10 != 0]
valid_data = [data[j] for i, j in enumerate(random_order) if i % 10 == 0]
def seq_padding(X, padding=0):
L = [len(x) for x in X]
class SemanticModel():
def __init__(self, batch_size=32, train=False):
self.batch_size = batch_size
if train:
chars = set()
train_datas = read_datas(TRAIN_DATA_FILE)
dev_datas = read_datas(DEV_DATA_FILE)
test_datas = read_datas(TEST_DATA_FILE)
for text1, text2, label in itertools.chain(train_datas, dev_datas):
chars.update(set(text1))
chars.update(set(text2))
_token_dict = load_vocab(dict_path) # 读取词典
self.token_dict, self.keep_words = {}, []
for c in ['[PAD]', '[UNK]', '[CLS]', '[SEP]', '[unused1]']:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
for c in chars:
if c in _token_dict:
self.token_dict[c] = len(self.token_dict)
self.keep_words.append(_token_dict[c])
self.tokenizer = SimpleTokenizer(self.token_dict) # 建立分词器
with open(os.path.join(model_save_path, 'tokenizer.pkl'),
"wb") as f:
pickle.dump(self.tokenizer, f)
with open(os.path.join(model_save_path, 'keep_words.pkl'),
"wb") as f:
pickle.dump(self.keep_words, f)
else:
with open(os.path.join(model_save_path, 'tokenizer.pkl'),
"rb") as f:
self.tokenizer = pickle.load(f)
with open(os.path.join(model_save_path, 'keep_words.pkl'),
"rb") as f:
self.keep_words = pickle.load(f)
self.model = self.make_model()
def make_model(self):
model = load_pretrained_model(config_path,
checkpoint_path,
keep_words=self.keep_words,
albert=True)
output = Lambda(lambda x: x[:, 0])(model.output)
# print(output.shape)
output = Dense(1,
activation='sigmoid')(output) # tanh, sigmoid, softmax
model = Model(inputs=model.input, outputs=output)
model.compile(
loss=
'binary_crossentropy', # categorical_crossentropy binary_crossentropy
optimizer=Adam(2e-6), # 用足够小的学习率
# optimizer=PiecewiseLinearLearningRate(Adam(1e-5), {1000: 1e-5, 2000: 6e-5}),
metrics=['accuracy'])
model.summary()
return model
def gnerator_data(self, file_name):
X1, X2, Y = [], [], []
while True:
for text1, text2, label in read_datas(file_name):
text1 = text1[:INPUT_LENGTH]
text2 = text2[:INPUT_LENGTH]
text1 = unicodedata.normalize('NFKD', text1).strip().lower()
text2 = unicodedata.normalize('NFKD', text2).strip().lower()
x1, x2 = self.tokenizer.encode(first=text1, second=text2)
y = int(label)
X1.append(x1)
X2.append(x2)
Y.append([y])
# Y.append(to_categorical(y))
if len(X1) == self.batch_size:
X1 = seq_padding(X1)
X2 = seq_padding(X2)
Y = seq_padding(Y)
# print(X1.shape, X2.shape, Y.shape)
yield [X1, X2], Y
X1, X2, Y = [], [], []
def train(self):
early_stopping = EarlyStopping(monitor='val_loss', patience=3)
model_checkpoint = ModelCheckpoint(filepath=os.path.join(
model_save_path,
'similarity-{epoch:02d}-{val_loss:.2f}-{val_acc:.3f}.hdf5'),
save_best_only=True,
save_weights_only=False)
tb = TensorBoard(
log_dir=log_dir, # log 目录
histogram_freq=0, # 按照何等频率（epoch）来计算直方图，0为不计算
batch_size=32, # 用多大量的数据计算直方图
write_graph=True, # 是否存储网络结构图
write_grads=False, # 是否可视化梯度直方图
write_images=False, # 是否可视化参数
embeddings_freq=0,
embeddings_layer_names=None,
embeddings_metadata=None)
hist = self.model.fit_generator(
self.gnerator_data(TRAIN_DATA_FILE),
steps_per_epoch=1000,
epochs=100,
validation_data=self.gnerator_data(DEV_DATA_FILE),
validation_steps=100,
callbacks=[early_stopping, model_checkpoint, tb])
print(hist.history.items())
def predict(self,
text1,
text2,
weitht_file='similarity-01-0.55-0.741.hdf5'):
self.model.load_weights(os.path.join(model_save_path, weitht_file),
by_name=True,
skip_mismatch=True,
reshape=True)
text1 = text1[:INPUT_LENGTH]
text2 = text2[:INPUT_LENGTH]
text1 = unicodedata.normalize('NFKD', text1).strip().lower()
text2 = unicodedata.normalize('NFKD', text2).strip().lower()
x1, x2 = self.tokenizer.encode(first=text1, second=text2)
X1 = seq_padding([x1])
X2 = seq_padding([x2])
ret = self.model.predict([X1, X2])
return ret
def batch_predict(self, question, database):
text1 = question
text1 = text1[:INPUT_LENGTH]
X1, X2 = [], []
for text2 in database:
text2 = text2[:INPUT_LENGTH]
text1 = unicodedata.normalize('NFKD', text1).strip().lower()
text2 = unicodedata.normalize('NFKD', text2).strip().lower()
x1, x2 = self.tokenizer.encode(first=text1, second=text2)
X1.append(x1)
X2.append(x2)
X1 = seq_padding(X1)
X2 = seq_padding(X2)
ret = self.model.predict([X1, X2])
return ret
import os
from bert4keras.bert import load_pretrained_model
from bert4keras.utils import SimpleTokenizer, load_vocab
import numpy as np
albert_model_path = '/home/gswyhq/github_projects/albert_zh/albert_large_zh'
# albert_model_path = '/notebooks/albert_zh/albert_large_zh'
# https://storage.googleapis.com/albert_zh/albert_large_zh.zip
config_path = os.path.join(albert_model_path, 'albert_config_large.json')
checkpoint_path = os.path.join(albert_model_path, 'albert_model.ckpt')
dict_path = os.path.join(albert_model_path, 'vocab.txt')
token_dict = load_vocab(dict_path) # 读取词典
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
model = load_pretrained_model(config_path, checkpoint_path,
with_mlm=True) # 建立模型，加载权重
# token_ids, segment_ids = tokenizer.encode(u'科学技术是第一生产力')
token_ids, segment_ids = tokenizer.encode(u'中国的首都是北京')
print('token_ids: {}, segment_ids: {}'.format(token_ids, segment_ids))
# mask掉“技术”
# token_ids[3] = token_ids[4] = token_dict['[MASK]']
token_ids[4] = token_ids[5] = token_dict['[MASK]']
# 用mlm模型预测被mask掉的部分
probas = model.predict([np.array([token_ids]), np.array([segment_ids])])[0]
# print(tokenizer.decode(probas[3:5].argmax(axis=1))) # 结果正是“技术”
class AlbertClassify:
def __init__(self,
initial_model=True,
model_path=os.path.join(CONFIG['model_dir'], 'albert.h5')):
self.initial_model = initial_model
token_dict = load_vocab(DICT_PATH)
self.tokenizer = SimpleTokenizer(token_dict)
self.model_path = model_path
if initial_model:
self.albert_model = load_pretrained_model(
CONFIG_PATH,
CHECKPOINT_PATH,
# keep_words=keep_words,
albert=True)
else:
self.load(model_path)
for l in self.albert_model.layers:
l.trainable = True
def train(self, train_data, valid_data):
train_D = DataGenerator(train_data, self.tokenizer,
CONFIG['batch_size'], CONFIG['max_len'])
valid_D = DataGenerator(valid_data, self.tokenizer,
CONFIG['batch_size'], CONFIG['max_len'])
output = Lambda(lambda x: x[:, 0])(self.albert_model.output)
output = Dense(1, activation='sigmoid')(output)
self.model = Model(self.albert_model.input, output)
save = ModelCheckpoint(os.path.join(self.model_path),
monitor='val_acc',
verbose=1,
save_best_only=True,
mode='auto')
early_stopping = EarlyStopping(monitor='val_acc',
min_delta=0,
patience=3,
verbose=1,
mode='auto')
callbacks = [save, early_stopping]
if self.initial_model:
x1_in = Input(shape=(None, ))
x2_in = Input(shape=(None, ))
x_in = self.albert_model([x1_in, x2_in])
x_in = Lambda(lambda x: x[:, 0])(x_in)
p = Dense(1, activation='sigmoid')(x_in)
self.model = Model([x1_in, x2_in], p)
else:
self.model = self.albert_model
self.model.compile(
loss='binary_crossentropy',
# optimizer=RAdam(1e-5), # 用足够小的学习率
optimizer=PiecewiseLinearLearningRate(Adam(1e-5), {
1000: 1e-5,
2000: 6e-5
}),
metrics=[
'accuracy', process.get_precision, process.get_recall,
process.get_f1
])
self.model.summary()
self.model.fit_generator(
train_D.__iter__(),
steps_per_epoch=len(train_D),
epochs=CONFIG['epochs'],
validation_data=valid_D.__iter__(),
validation_steps=len(valid_D),
callbacks=callbacks,
use_multiprocessing=CONFIG['use_multiprocessing'],
)
def predict(self, test_data):
"""
预测
:param test_data:
:return:
"""
X1 = []
X2 = []
for s in test_data:
x1, x2 = self.tokenizer.encode(first=s[:CONFIG['max_len']])
X1.append(x1)
X2.append(x2)
X1 = seq_padding(X1)
X2 = seq_padding(X2)
predict_results = self.model.predict([X1, X2])
return predict_results
def load(self, model_path):
"""
load the pre-trained model
"""
try:
self.albert_model = load_model(str(model_path),
custom_objects=get_custom_objects(),
compile=False)
except Exception as ex:
print('load error')
return self