def build_model(self):
import tensorflow as tf
from keras.backend.tensorflow_backend import set_session
config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC' # A "Best-fit with coalescing" algorithm, simplified from a version of dlmalloc.
if self.memory_fraction:
config.gpu_options.per_process_gpu_memory_fraction = self.memory_fraction
config.gpu_options.allow_growth = False
else:
config.gpu_options.allow_growth = True
set_session(tf.Session(config=config))
# 补充输入
subject_labels = Input(shape=(None, 2), name='Subject-Labels')
subject_ids = Input(shape=(2, ), name='Subject-Ids')
object_labels = Input(shape=(None, self.num_classes, 2),
name='Object-Labels')
# 加载预训练模型
bert = build_transformer_model(
config_path=self.bert_config_path,
checkpoint_path=self.bert_checkpoint_path,
return_keras_model=False,
)
# 预测subject
output = Dense(units=2,
activation='sigmoid',
kernel_initializer=bert.initializer)(bert.model.output)
subject_preds = Lambda(lambda x: x**2)(output)
self.subject_model = Model(bert.model.inputs, subject_preds)
# 传入subject，预测object
# 通过Conditional Layer Normalization将subject融入到object的预测中
output = bert.model.layers[-2].get_output_at(-1)
subject = Lambda(self.extrac_subject)([output, subject_ids])
output = LayerNormalization(conditional=True)([output, subject])
output = Dense(units=self.num_classes * 2,
activation='sigmoid',
kernel_initializer=bert.initializer)(output)
output = Lambda(lambda x: x**4)(output)
object_preds = Reshape((-1, self.num_classes, 2))(output)
self.object_model = Model(bert.model.inputs + [subject_ids],
object_preds)
# 训练模型
self.model = Model(
bert.model.inputs + [subject_labels, subject_ids, object_labels],
[subject_preds, object_preds])
mask = bert.model.get_layer('Embedding-Token').output_mask
mask = K.cast(mask, K.floatx())
subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
subject_loss = K.mean(subject_loss, 2)
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)
object_loss = K.binary_crossentropy(object_labels, object_preds)
object_loss = K.sum(K.mean(object_loss, 3), 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)
self.model.add_loss(subject_loss + object_loss)
AdamEMA = extend_with_exponential_moving_average(Adam, name='AdamEMA')
self.optimizer = AdamEMA(lr=1e-4)
def build_model():
bert_model = build_transformer_model(
config_path=Config.config_path,
checkpoint_path=Config.checkpoint_path,
return_keras_model=False)
# 补充输入
subject_labels = Input(shape=(None, 2))
subject_ids = Input(shape=(2, ))
object_labels = Input(shape=(None, len(predicate2id), 2))
# 预测subject
output = Dense(units=2,
activation='sigmoid',
kernel_initializer=bert_model.initializer)(
bert_model.model.output)
subject_preds = Lambda(lambda x: x**2)(output)
subject_model = Model(bert_model.inputs, subject_preds)
# 传入subject，预测object
output = bert_model.model.layers[-2].get_output_at(-1)
subject = Lambda(extrac_subject)([output, subject_ids])
output = LayerNormalization(conditional=True)([output, subject])
output = Dense(units=len(predicate2id) * 2,
activation='sigmoid',
kernel_initializer=bert_model.initializer)(output)
output = Lambda(lambda x: x**4)(output)
object_preds = Reshape((-1, len(predicate2id), 2))(output)
object_model = Model(bert_model.model.inputs + [subject_ids], object_preds)
# 训练模型
train_model = Model(
bert_model.model.inputs + [subject_labels, subject_ids, object_labels],
[subject_preds, object_preds])
mask = bert_model.model.get_layer('Embedding-Token').output_mask
mask = K.cast(mask, K.floatx())
subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
subject_loss = K.mean(subject_loss, 2)
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)
object_loss = K.binary_crossentropy(object_labels, object_preds)
object_loss = K.sum(K.mean(object_loss, 3), 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)
train_model.add_loss(subject_loss + object_loss)
optimizer = Adam(Config.learning_rate)
train_model.compile(optimizer=optimizer)
return train_model, subject_model, object_model
def compute_loss(self, inputs, mask=None):
subject_labels, object_labels = inputs[:2]
subject_preds, object_preds, _ = inputs[2:]
if mask[4] is None:
mask = 1.0
else:
mask = K.cast(mask[4], K.floatx())
# subject部分loss
subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
subject_loss = K.mean(subject_loss, 2)
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)
# object部分loss
object_loss = K.binary_crossentropy(object_labels, object_preds)
object_loss = K.sum(K.mean(object_loss, 3), 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)
# 总的loss
return subject_loss + object_loss
def compute_loss(self, inputs, mask=None):
subject_labels, object_labels = inputs[:2]
subject_preds, object_preds, _ = inputs[2:]
if mask[4] is None:
mask = 1.0
else:
mask = K.cast(mask[4], K.floatx())
subject_loss = K.binary_crossentropy(
subject_labels, subject_preds) # (btz, seq_len, 2)在最后一维上计算loss
subject_loss = K.mean(
subject_loss, 2) # (btz, seq_len)就像是(btz, seq_len, 1) 每个字的平均loss
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)
object_loss = K.binary_crossentropy(object_labels, object_preds)
object_loss = K.sum(K.mean(object_loss, 3), 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)
return subject_loss + object_loss
# activation='sigmoid',
# kernel_initializer=bert.initializer)(output)
# output = Reshape((-1, len(predicate2id), 2))(output) #[? ? predicate49*2]->[? ? 49 2]
# object_preds = Lambda(lambda x: x**4)(output)
#
# object_model = Model(bert.model.inputs + [subject_ids], object_preds) #sub,text -> obj,predicate
# 训练模型
# train_model = Model(bert.model.inputs + [subject_labels,
# subject_ids, object_labels],
# [subject_preds, object_preds])
train_model = Model(bert.model.inputs + [subject_labels], [subject_preds])
mask = bert.model.get_layer('Sequence-Mask').output_mask
subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
subject_loss = K.sum(K.mean(subject_loss, 3), 2)
subject_loss = K.sum(subject_loss * mask) / K.sum(mask)
#
# object_loss = K.binary_crossentropy(object_labels, object_preds)
# object_loss = K.sum(K.mean(object_loss, 3), 2)
# object_loss = K.sum(object_loss * mask) / K.sum(mask)
train_model.add_loss(subject_loss)
train_model.compile(optimizer=Adam(1e-5))
#
# def extract_spoes(text):
# #抽取输入text所包含的三元组
#
def build_model():
"""
调用模型参数，搭建事件抽取模型主体，先搭建触发词模型，然后围绕触发词下标搭建其他论元模型。
:return: 各个论元模型对象
"""
with SESS.as_default():
with SESS.graph.as_default():
# 构建bert模型主体
bert_model = build_transformer_model(
config_path=bert_config.config_path,
checkpoint_path=bert_config.checkpoint_path,
return_keras_model=False,
model=bert_config.model_type)
# l为模型内部的层名，格式为--str
for l in bert_model.layers:
bert_model.model.get_layer(l).trainable = True
# 搭建模型
# keras会自动对所有的占位张量添加batch_size维度
# 动词输入 (batch_size, seq_len)
trigger_start_in = Input(shape=(None, ))
trigger_end_in = Input(shape=(None, ))
# 动词下标输入 (batch_size, seq_len)
trigger_index_start_in = Input(shape=(1, ))
trigger_index_end_in = Input(shape=(1, ))
# 宾语输入 (batch_size, seq_len)
object_start_in = Input(shape=(None, ))
object_end_in = Input(shape=(None, ))
# 主语输入 (batch_size, seq_len)
subject_start_in = Input(shape=(None, ))
subject_end_in = Input(shape=(None, ))
# 地点输入 (batch_size, seq_len)
loc_start_in = Input(shape=(None, ))
loc_end_in = Input(shape=(None, ))
# 时间输入 (batch_size, seq_len)
time_start_in = Input(shape=(None, ))
time_end_in = Input(shape=(None, ))
# 否定词输入 (batch_size, seq_len)
negative_start_in = Input(shape=(None, ))
negative_end_in = Input(shape=(None, ))
# 将输入的占位符赋值给相应的变量(此处只是在使用时方便，没有其他的模型结构意义)
# 动词输入
trigger_start, trigger_end = trigger_start_in, trigger_end_in
# 动词下标
trigger_index_start, trigger_index_end = trigger_index_start_in, trigger_index_end_in
# 宾语输入
object_start, object_end = object_start_in, object_end_in
# 主语输入
subject_start, subject_end = subject_start_in, subject_end_in
# 地点输入
loc_start, loc_end = loc_start_in, loc_end_in
# 时间输入
time_start, time_end = time_start_in, time_end_in
# 否定词输入
negative_start, negative_end = negative_start_in, negative_end_in
# bert_model.model.inputs为列表格式，含有两个张量，[token_ids(batch, seq_len), segment_ids(batch, seq_len)]
# mask操作，将bert模型的输入的token_ids序列，进行维度扩充[batch_size, seq_len, 1],
# 然后将初始填充为0的地方全部都用0代替，非0的地方都用1占位，
# 这是为后边计算损失做准备，防止计算损失时，前期填充为0的位置也进行反向传播
mask = Lambda(lambda x: K.cast(
K.greater(K.expand_dims(x[0], 2), 0), 'float32'))(
bert_model.model.inputs)
# 计算动词输出的起始终止标签，bert_model.model.output [batch_size, seq_len, 768]
trigger_start_out = Dense(1, activation='sigmoid')(
bert_model.model.output)
trigger_end_out = Dense(1, activation='sigmoid')(
bert_model.model.output)
# 预测trigger动词的模型
trigger_model = Model(bert_model.model.inputs,
[trigger_start_out, trigger_end_out])
# 将动词下标对应位置的子向量抽取出来并计算均值
k1v = Lambda(seq_gather)(
[bert_model.model.output, trigger_index_start])
k2v = Lambda(seq_gather)(
[bert_model.model.output, trigger_index_end])
kv = Average()([k1v, k2v])
# 融合动词词向量的句子张量，用来作为预测其它论元部分的向量
t = LayerNormalization(conditional=True)(
[bert_model.model.output, kv])
# 宾语模型输出
object_start_out = Dense(1, activation='sigmoid')(t)
object_end_out = Dense(1, activation='sigmoid')(t)
# 主语模型输出
subject_start_out = Dense(1, activation='sigmoid')(t)
subject_end_out = Dense(1, activation='sigmoid')(t)
# 地点模型输出
loc_start_out = Dense(1, activation='sigmoid')(t)
loc_end_out = Dense(1, activation='sigmoid')(t)
# 时间模型输出
time_start_out = Dense(1, activation='sigmoid')(t)
time_end_out = Dense(1, activation='sigmoid')(t)
# 否定词模型输出
negative_start_out = Dense(1, activation='sigmoid')(t)
negative_end_out = Dense(1, activation='sigmoid')(t)
# 输入text和trigger，预测object
object_model = Model(
bert_model.model.inputs +
[trigger_index_start_in, trigger_index_end_in],
[object_start_out, object_end_out])
# 输入text和trigger，预测subject
subject_model = Model(
bert_model.model.inputs +
[trigger_index_start_in, trigger_index_end_in],
[subject_start_out, subject_end_out])
# 输入text和trigger，预测loc
loc_model = Model(
bert_model.model.inputs +
[trigger_index_start_in, trigger_index_end_in],
[loc_start_out, loc_end_out])
# 输入text和trigger，预测time
time_model = Model(
bert_model.model.inputs +
[trigger_index_start_in, trigger_index_end_in],
[time_start_out, time_end_out])
# 否定词模型
negative_model = Model(
bert_model.model.inputs +
[trigger_index_start_in, trigger_index_end_in],
[negative_start_out, negative_end_out])
# 主模型
train_model = Model(
bert_model.model.inputs + [
trigger_start_in, trigger_end_in, trigger_index_start_in,
trigger_index_end_in, object_start_in, object_end_in,
subject_start_in, subject_end_in, loc_start_in, loc_end_in,
time_start_in, time_end_in, negative_start_in,
negative_end_in
], [
trigger_start_out, trigger_end_out, object_start_out,
object_end_out, subject_start_out, subject_end_out,
loc_start_out, loc_end_out, time_start_out, time_end_out,
negative_start_out, negative_end_out
])
# 扩充维度, 构造成与mask矩阵相同的结构，方便后续计算模型各部分损失
trigger_start = K.expand_dims(trigger_start, 2)
trigger_end = K.expand_dims(trigger_end, 2)
object_start = K.expand_dims(object_start, 2)
object_end = K.expand_dims(object_end, 2)
subject_start = K.expand_dims(subject_start, 2)
subject_end = K.expand_dims(subject_end, 2)
loc_start = K.expand_dims(loc_start, 2)
loc_end = K.expand_dims(loc_end, 2)
time_start = K.expand_dims(time_start, 2)
time_end = K.expand_dims(time_end, 2)
negative_start = K.expand_dims(negative_start, 2)
negative_end = K.expand_dims(negative_end, 2)
# 构造模型损失函数，使用mask矩阵将前期填充为0的位置全部掩掉，不进行反向传播。
# 动词损失
trigger_start_loss = K.binary_crossentropy(trigger_start,
trigger_start_out)
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
trigger_start_loss = K.sum(trigger_start_loss * mask) / K.sum(mask)
trigger_end_loss = K.binary_crossentropy(trigger_end,
trigger_end_out)
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
trigger_end_loss = K.sum(trigger_end_loss * mask) / K.sum(mask)
# 宾语损失
object_start_loss = K.sum(
K.binary_crossentropy(object_start, object_start_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
object_start_loss = K.sum(object_start_loss * mask) / K.sum(mask)
object_end_loss = K.sum(
K.binary_crossentropy(object_end, object_end_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
object_end_loss = K.sum(object_end_loss * mask) / K.sum(mask)
# 主语损失
subject_start_loss = K.sum(
K.binary_crossentropy(subject_start, subject_start_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
subject_start_loss = K.sum(subject_start_loss * mask) / K.sum(mask)
subject_end_loss = K.sum(
K.binary_crossentropy(subject_end, subject_end_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
subject_end_loss = K.sum(subject_end_loss * mask) / K.sum(mask)
# 地点损失
loc_start_loss = K.sum(
K.binary_crossentropy(loc_start, loc_start_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
loc_start_loss = K.sum(loc_start_loss * mask) / K.sum(mask)
loc_end_loss = K.sum(K.binary_crossentropy(loc_end, loc_end_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
loc_end_loss = K.sum(loc_end_loss * mask) / K.sum(mask)
# 时间损失
time_start_loss = K.sum(
K.binary_crossentropy(time_start, time_start_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
time_start_loss = K.sum(time_start_loss * mask) / K.sum(mask)
time_end_loss = K.sum(K.binary_crossentropy(
time_end, time_end_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
time_end_loss = K.sum(time_end_loss * mask) / K.sum(mask)
# 否定词损失
negative_start_loss = K.sum(
K.binary_crossentropy(negative_start, negative_start_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
negative_start_loss = K.sum(
negative_start_loss * mask) / K.sum(mask)
negative_end_loss = K.sum(
K.binary_crossentropy(negative_end, negative_end_out))
# 使用mask矩阵，将前期填充为0的位置掩掉，不进行反向传播
negative_end_loss = K.sum(negative_end_loss * mask) / K.sum(mask)
# 合并损失
loss = (trigger_start_loss + trigger_end_loss) + (
object_start_loss +
object_end_loss) + (subject_start_loss + subject_end_loss) + (
loc_start_loss +
loc_end_loss) + (time_start_loss + time_end_loss) + (
negative_start_loss + negative_end_loss)
train_model.add_loss(loss)
train_model.compile(
optimizer=Adam(extract_train_config.learning_rate))
train_model.summary()
return trigger_model, subject_model, object_model, time_model, loc_model, negative_model, train_model
x = bert_model.get_layer(output_layers).output
start_output = Dense(2,activation='sigmoid',name='start')(x)
end_output = Dense(2,activation='sigmoid',name='end')(x)
start_output = Lambda(lambda x:x ** 2)(start_output)
end_output = Lambda(lambda x:x ** 2)(end_output)
start_model = Model(bert_model.input,start_output)
end_model = Model(bert_model.input,end_output)
model = Model(bert_model.input + [start_labels,end_labels],[start_output,end_output])
model.summary()
start_loss = K.binary_crossentropy(start_labels,start_output)
start_loss = K.mean(start_loss,2)
start_loss = K.sum(start_loss * mask) / K.sum(mask)
end_loss = K.binary_crossentropy(end_labels,end_output)
end_loss = K.mean(end_loss,2)
end_loss = K.sum(end_loss * mask) / K.sum(mask)
loss = start_loss + end_loss
model.add_loss(loss)
model.compile(optimizer=Adam(learning_rate))
def extract(qtext):
v = qtext.split('fengefu')[0]
#object_model = Model(bert.model.inputs + [subject_ids], object_preds) #sub,text -> obj,predicate
# 训练模型
# train_model = Model(bert.model.inputs + [subject_labels, subject_ids, object_labels],
# [subject_preds, object_preds])
train_model = Model(
bert.model.inputs + [subject_ids, predicate_id, object_labels],
[object_preds])
mask = bert.model.get_layer('Sequence-Mask').output_mask
# subject_loss = K.binary_crossentropy(subject_labels, subject_preds)
# subject_loss = K.mean(subject_loss, 2)
# subject_loss = K.sum(subject_loss * mask) / K.sum(mask)
object_loss = K.binary_crossentropy(object_labels,
object_preds) # [batch step 2]
object_loss = K.mean(object_loss, 2)
object_loss = K.sum(object_loss * mask) / K.sum(mask)
train_model.add_loss(object_loss)
train_model.compile(optimizer=Adam(1e-5))
def extract_spoes(text):
#抽取输入text所包含的三元组
tokens = tokenizer.tokenize(text, max_length=maxlen)
token_ids, segment_ids = tokenizer.encode(text, max_length=maxlen)
# 抽取subject
subject_preds = subject_model.predict([[token_ids], [segment_ids]])
start = np.where(subject_preds[0, :, 0] > 0.6)[0]
