def build_model(config_path,
checkpoint_path,
max_seq_length,
bert_trainable=False):
in_id = Input(shape=(max_seq_length, ), name="input_ids", dtype="int32")
in_segment = Input(shape=(max_seq_length, ),
name="segment_ids",
dtype="int32")
bert_model = load_pretrained_model(config_path,
checkpoint_path) # 建立模型，加载权重
for l in bert_model.layers:
if bert_trainable:
l.trainable = True
else:
l.trainable = False
sequence_output = bert_model([in_id, in_segment])
pooled_output = Lambda(lambda x: x[:, 0])(sequence_output)
pred = Dense(1, activation="sigmoid")(pooled_output)
model = Model(inputs=[in_id, in_segment], outputs=pred)
model.compile(loss="binary_crossentropy",
optimizer=Adam(lr=0.001),
metrics=["accuracy"])
model.summary(line_length=150)
return model
def build_model(config_path,
checkpoint_path,
max_seq_length,
label_num,
bert_trainable=False):
in_id = Input(shape=(max_seq_length, ), name="input_ids", dtype="int32")
in_segment = Input(shape=(max_seq_length, ),
name="segment_ids",
dtype="int32")
bert_model = load_pretrained_model(config_path,
checkpoint_path) # 建立模型，加载权重
for l in bert_model.layers:
if bert_trainable:
l.trainable = True
else:
l.trainable = False
sequence_output = bert_model([in_id, in_segment])
bilstm_output = Bidirectional(CuDNNLSTM(
128, return_sequences=True))(sequence_output)
layer_dense = Dense(64, activation='tanh', name='layer_dense')
layer_crf_dense = Dense(label_num, name='layer_crf_dense')
layer_crf = CRF(label_num, name='layer_crf')
dense = layer_dense(bilstm_output)
dense = layer_crf_dense(dense)
pred = layer_crf(dense)
model = Model(inputs=[in_id, in_segment], outputs=pred)
model.compile(loss=layer_crf.loss,
optimizer=Adam(lr=1e-5),
metrics=[layer_crf.viterbi_accuracy])
model.summary(line_length=150)
return model
K.set_session(session)
config_path = "/home/new/Toxicity/bert_model/models/chinese_L-12_H-768_A-12/bert_config.json"
checkpoint_path = "/home/new/Toxicity/bert_model/models/chinese_L-12_H-768_A-12/bert_model.ckpt"
vocab_file = "/home/new/Toxicity/bert_model/models/chinese_L-12_H-768_A-12/vocab.txt"
max_seq_length = 128
maxlen = 100
config_path = '/home/new/Toxicity/bert_model/models/chinese_L-12_H-768_A-12/bert_config.json'
checkpoint_path = '/home/new/Toxicity/bert_model/models/chinese_L-12_H-768_A-12/bert_model.ckpt'
dict_path = '/home/new/Toxicity/bert_model/models/chinese_L-12_H-768_A-12/vocab.txt'
token_dict = load_vocab(dict_path) # 读取词典
tokenizer = SimpleTokenizer(token_dict) # 建立分词器
bert_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(bert_model.predict([np.array([token_ids]), np.array([segment_ids])]).shape)
# save_h5_model(model)
# tf.keras.experimental.export_saved_model(model, saved_model_path)
"""
train examles:
0 选择珠江花园的原因就是方便，有电动扶梯直接到达海边，周围餐馆、食廊、商场、超市、摊位一应俱全。酒店装修一般，但还算整洁。 泳池在大堂的屋顶，因此很小，不过女儿倒是喜欢。 包的早餐是西式的，还算丰富。 服务吗，一般 None 1
1 15.4寸笔记本的键盘确实爽，基本跟台式机差不多了，蛮喜欢数字小键盘，输数字特方便，样子也很美观，做工也相当不错 None 1
2 房间太小。其他的都一般。。。。。。。。。 None 0
3 1.接电源没有几分钟,电源适配器热的不行. 2.摄像头用不起来. 3.机盖的钢琴漆，手不能摸，一摸一个印. 4.硬盘分区不好办. None 0