def build_model(self):
with tf.variable_scope("knowledge-embedding-layer"):
knowledge_embedding = tf.Variable(tf.random_normal([2, self.config.knowledge_dimension],
stddev=0, seed=1), trainable=True,
name='knowledge-embedding')
self.x2_label_embedding = tf.nn.embedding_lookup(knowledge_embedding, self.x2_label)
with tf.variable_scope("first-CNN-layer"):
self.output_x1_1 = tf.layers.conv1d(self.input_X1,filters=self.config.filters_num,kernel_size=self.config.first_kernel_size,padding='same',name='first-cnn1')
# self.output_x1_1 = tf.nn.dropout(self.output_x1_1, self.dropout_rate)
self.output_x2_1 = tf.layers.conv1d(self.input_X2,filters=self.config.filters_num,kernel_size=self.config.first_kernel_size,padding='same',name='first-cnn2')
# self.output_x2_1 = tf.nn.dropout(self.output_x2_1, self.dropout_rate)
with tf.variable_scope("first-interaction"):
interaction = modules.Interaction(10, self.output_x1_1,self.output_x2_1,self.x2_label_embedding)
self.inter1_output_x2 = interaction.exeInteraction()
with tf.variable_scope("second-CNN-layer"):
self.output_x1_2 = tf.layers.conv1d(self.output_x1_1,filters=self.config.filters_num,kernel_size=self.config.second_kernel_size,padding='same',name='second-cnn1')
self.output_x2_2 = tf.layers.conv1d(self.output_x2_1,filters=self.config.filters_num,kernel_size=self.config.second_kernel_size,padding='same',name='second-cnn2')
# self.output_x1_2 = tf.nn.dropout(self.output_x1_2, self.dropout_rate)
# self.output_x2_2 = tf.nn.dropout(self.output_x2_2, self.dropout_rate)
with tf.variable_scope("second-interaction"):
interaction = modules.Interaction(10, self.output_x1_2, self.output_x2_2,self.x2_label_embedding)
self.inter2_output_x2 = interaction.exeInteraction()
with tf.variable_scope("third-CNN-layer"):
self.output_x1_3 = tf.layers.conv1d(self.output_x1_2,filters=self.config.filters_num,kernel_size=self.config.third_kernel_size,padding='same',name='third-cnn1')
self.output_x2_3 = tf.layers.conv1d(self.output_x2_2,filters=self.config.filters_num,kernel_size=self.config.third_kernel_size,padding='same',name='third-cnn2')
# self.output_x1_3 = tf.nn.dropout(self.output_x1_3, self.dropout_rate)
# self.output_x2_3 = tf.nn.dropout(self.output_x2_3, self.dropout_rate)
with tf.variable_scope("third-interaction"):
interaction = modules.Interaction(10, self.output_x1_3,self.output_x2_3,self.x2_label_embedding)
self.inter3_output_x2 = interaction.exeInteraction()
with tf.variable_scope("fusion-layer"):
self.fusion_output = tf.concat([self.inter1_output_x2, self.inter2_output_x2,self.inter3_output_x2], axis=-1) # [Batch, 3 * len]
with tf.variable_scope("predict-layer"):
self.output_ = tf.nn.relu(tf.layers.dense(inputs=self.fusion_output,units=self.config.mlp_output,name='fnn1'))
self.output_ = tf.layers.dropout(self.output_,rate=self.dropout_rate)
self.logit = tf.layers.dense(inputs=self.output_,units=2,name='fnn2')
with tf.variable_scope("optimize-layer"):
self.pred_y = tf.argmax(tf.nn.softmax(self.logit), 1)
# 损失函数，交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logit,
labels=self.y) # 对logits进行softmax操作后，做交叉墒，输出的是一个向量
self.loss = tf.reduce_mean(cross_entropy) # 将交叉熵向量求和，即可得到交叉熵
# 优化器
self.optim = tf.train.AdamOptimizer(learning_rate=param.BaseConfig.lr).minimize(self.loss)
with tf.name_scope("accuracy"):
# 准确率
correct_pred = tf.equal(tf.argmax(self.y, 1),
self.pred_y) # 由于input_y也是onehot编码，因此，调用tf.argmax(self.input_y)得到的是1所在的下表
self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
def build_model(self):
#WIL-1 word-Interaction-layer
# with tf.variable_scope("word-Interaction-layer"):
# interaction = modules.Interaction(4, self.input_X1, self.input_X2)
# self.inter0_output_x2 = interaction.exeInteraction()
#WIL-2 word-Interaction-layer
with tf.variable_scope("word-Interaction-layer"):
self.beta = tf.Variable(tf.random_normal(shape=[3,1], stddev=0, seed=1, dtype=tf.float32), trainable=True, name='beta')
interaction = modules.Interaction(6, self.input_X1, self.input_X2, self.x2_label, self.beta)
self.inter0_output_x2 = interaction.exeInteraction()
#添加一个mean pooling
# self.inter0_output_x2 = tf.expand_dims(tf.expand_dims(self.inter0_output_x2,axis=2),axis=3)
# self.inter0_output_x2 = tf.nn.avg_pool(self.inter0_output_x2,ksize=[1,3,1,1],strides=[1,1,1,1],padding='VALID',name='mean-pooling')
with tf.variable_scope("first-CNN-layer"):
self.output_x1_1 = tf.layers.conv1d(self.input_X1,filters=self.config.filters_num,kernel_size=self.config.first_kernel_size,padding='same',name='first-cnn1')
self.output_x2_1 = tf.layers.conv1d(self.input_X2,filters=self.config.filters_num,kernel_size=self.config.first_kernel_size,padding='same',name='first-cnn2')
with tf.variable_scope("first-interaction"):
self.beta1 = tf.Variable(tf.random_normal(shape=[3,1], stddev=0, seed=1, dtype=tf.float32), trainable=True,
name='beta1')
interaction = modules.Interaction(6, self.output_x1_1,self.output_x2_1,self.x2_label,self.beta1)
self.inter1_output_x2 = interaction.exeInteraction()
# self.inter1_output_x2 = self.interaction(self.output_x1_1, self.output_x2_1)
#mean pooling
# self.inter1_output_x2 = tf.expand_dims(tf.expand_dims(self.inter1_output_x2,axis=2),axis=3)
# self.inter1_output_x2 = tf.nn.avg_pool(self.inter1_output_x2,ksize=[1,3,1,1],strides=[1,1,1,1],padding='VALID', name='mean-pooling')
# self.inter1_output_x2 = tf.reshape(self.inter1_output_x2, shape=[-1, 28])
with tf.variable_scope("second-CNN-layer"):
self.output_x1_2 = tf.layers.conv1d(self.output_x1_1,filters=self.config.filters_num,kernel_size=self.config.second_kernel_size,padding='same',name='second-cnn1')
self.output_x2_2 = tf.layers.conv1d(self.output_x2_1,filters=self.config.filters_num,kernel_size=self.config.second_kernel_size,padding='same',name='second-cnn2')
with tf.variable_scope("second-interaction"):
self.beta2 = tf.Variable(tf.random_normal(shape=[3,1], stddev=0, seed=1, dtype=tf.float32), trainable=True,
name='beta2')
interaction = modules.Interaction(6, self.output_x1_2, self.output_x2_2,self.x2_label,self.beta2)
self.inter2_output_x2 = interaction.exeInteraction()
# self.inter2_output_x2 = self.interaction(self.output_x1_2, self.output_x2_2)
#mean pooling
# self.inter2_output_x2 = tf.expand_dims(tf.expand_dims(self.inter2_output_x2,axis=2),axis=3)
# self.inter2_output_x2 = tf.nn.avg_pool(self.inter2_output_x2,ksize=[1,3,1,1],strides=[1,1,1,1],padding='VALID', name='mean-pooling')
# self.inter2_output_x2 = tf.reshape(self.inter2_output_x2,shape=[-1,28])
with tf.variable_scope("third-CNN-layer"):
self.output_x1_3 = tf.layers.conv1d(self.output_x1_2,filters=self.config.filters_num,kernel_size=self.config.third_kernel_size,padding='same',name='third-cnn1')
self.output_x2_3 = tf.layers.conv1d(self.output_x2_2,filters=self.config.filters_num,kernel_size=self.config.third_kernel_size,padding='same',name='third-cnn2')
with tf.variable_scope("third-interaction"):
self.beta3 = tf.Variable(tf.random_normal(shape=[3,1], stddev=0, seed=1, dtype=tf.float32), trainable=True,
name='beta3')
interaction = modules.Interaction(6, self.output_x1_3,self.output_x2_3,self.x2_label,self.beta3)
self.inter3_output_x2 = interaction.exeInteraction()
# self.inter3_output_x2 = self.interaction(self.output_x1_3,self.output_x2_3)
#mean pooling
# self.inter3_output_x2 = tf.expand_dims(tf.expand_dims(self.inter3_output_x2,axis=2),axis=3)
# self.inter3_output_x2 = tf.nn.avg_pool(self.inter3_output_x2,ksize=[1,3,1,1],strides=[1,1,1,1],padding='VALID', name='mean-pooling')
# self.inter3_output_x2 = tf.reshape(self.inter3_output_x2, shape=[-1, 28])
with tf.variable_scope("fusion-layer"):
#v2
# self.fusion_output = tf.stack(
# [self.inter1_output_x2, self.inter2_output_x2, self.inter3_output_x2],
# axis=-1) # [Batch,len,3]
# self.fusion_output = tf.layers.conv1d(self.output_x1_1, filters=64,
# kernel_size=self.config.second_kernel_size, padding='same',
# name='fifth-cnn1')
# self.fusion_output = tf.reduce_max(self.fusion_output, axis=-1)
# #v3
# self.fusion_output = tf.stack([self.inter1_output_x2, self.inter2_output_x2,self.inter3_output_x2,self.x2_label], axis=-1) # [Batch,len,4]
# self.fusion_output = tf.layers.conv1d(self.output_x1_1,filters=64,kernel_size=self.config.second_kernel_size,padding='same',name='fifth-cnn1')
# self.fusion_output = tf.reduce_max(self.fusion_output,axis=-1)
#v4
# self.fusion_output = tf.concat([self.inter1_output_x2, self.inter2_output_x2,self.inter3_output_x2],axis=-1)
#v5
self.fusion_output = tf.concat([self.inter0_output_x2, self.inter1_output_x2, self.inter2_output_x2, self.inter3_output_x2],
axis=-1)
with tf.variable_scope("predict-layer"):
self.output_ = tf.nn.relu(tf.layers.dense(inputs=self.fusion_output,units=self.config.mlp_output,name='fnn1'))
self.output_ = tf.layers.dropout(self.output_,rate=self.config.dropout_rate)
self.logit = tf.layers.dense(inputs=self.output_,units=2,name='fnn2')
with tf.variable_scope("optimize-layer"):
self.pred_y = tf.argmax(tf.nn.softmax(self.logit), 1)
# 损失函数，交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=self.logit,
labels=self.y) # 对logits进行softmax操作后，做交叉墒，输出的是一个向量
self.loss = tf.reduce_mean(cross_entropy) # 将交叉熵向量求和，即可得到交叉熵
# 优化器
self.optim = tf.train.AdamOptimizer(learning_rate=param.BaseConfig.lr).minimize(self.loss)
with tf.name_scope("accuracy"):
# 准确率
correct_pred = tf.equal(tf.argmax(self.y, 1),
self.pred_y) # 由于input_y也是onehot编码，因此，调用tf.argmax(self.input_y)得到的是1所在的下表
self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
def build_model(self):
with tf.variable_scope("first-CNN-layer"):
self.output_x1_1 = tf.layers.conv1d(
self.input_X1,
filters=self.config.filters_num,
kernel_size=self.config.first_kernel_size,
padding='same',
name='first-cnn1')
self.output_x2_1 = tf.layers.conv1d(
self.input_X2,
filters=self.config.filters_num,
kernel_size=self.config.first_kernel_size,
padding='same',
name='first-cnn2')
self.pooling_x2_1 = tf.reduce_max(self.output_x2_1,
axis=-1) #[B,l]
with tf.variable_scope("second-CNN-layer"):
self.output_x1_2 = tf.layers.conv1d(
self.output_x1_1,
filters=self.config.filters_num,
kernel_size=self.config.second_kernel_size,
padding='same',
name='second-cnn1')
self.output_x2_2 = tf.layers.conv1d(
self.output_x2_1,
filters=self.config.filters_num,
kernel_size=self.config.second_kernel_size,
padding='same',
name='second-cnn2')
with tf.variable_scope("third-CNN-layer"):
self.output_x1_3 = tf.layers.conv1d(
self.output_x1_2,
filters=self.config.filters_num,
kernel_size=self.config.third_kernel_size,
padding='same',
name='third-cnn1')
self.output_x2_3 = tf.layers.conv1d(
self.output_x2_2,
filters=self.config.filters_num,
kernel_size=self.config.third_kernel_size,
padding='same',
name='third-cnn2')
self.pooling_x2_3 = tf.reduce_max(self.output_x2_3, axis=-1)
with tf.variable_scope("interaction"):
interaction11 = modules.Interaction(4, self.output_x1_1,
self.output_x2_1)
self.inter11 = interaction11.exeInteraction()
# interaction21 = modules.Interaction(4, self.output_x1_2, self.output_x2_1)
# self.inter21 = interaction21.exeInteraction()
# interaction31 = modules.Interaction(4, self.output_x1_3, self.output_x2_1)
# self.inter31 = interaction31.exeInteraction()
# interaction12 = modules.Interaction(4, self.output_x1_1,self.output_x2_2)
# self.inter12 = interaction12.exeInteraction()
interaction22 = modules.Interaction(4, self.output_x1_2,
self.output_x2_2)
self.inter22 = interaction22.exeInteraction()
# interaction32 = modules.Interaction(4, self.output_x1_3, self.output_x2_2)
# self.inter32 = interaction32.exeInteraction()
#
# interaction13 = modules.Interaction(4, self.output_x1_1,self.output_x2_3)
# self.inter13 = interaction13.exeInteraction()
#
# interaction23 = modules.Interaction(4, self.output_x1_2, self.output_x2_3)
# self.inter23 = interaction23.exeInteraction()
interaction33 = modules.Interaction(4, self.output_x1_3,
self.output_x2_3)
self.inter33 = interaction33.exeInteraction() #[B,l]
# self.inter = tf.reshape(tf.stack([self.inter11,self.inter12,self.inter13,self.inter12,self.inter22,self.inter32,self.inter31,self.inter32,self.inter33],axis=-1),
# shape=[-1,self.config.Y_maxlen,9]) #[B,9,l]
self.inter = tf.reshape(tf.stack(
[self.inter11, self.inter22, self.inter33], axis=-1),
shape=[-1, self.config.Y_maxlen, 3])
with tf.variable_scope("fusion-layer"):
self.fusion_output = tf.reduce_max(self.inter, axis=1)
with tf.variable_scope("predict-layer"):
self.output_ = tf.nn.relu(
tf.layers.dense(inputs=self.fusion_output,
units=self.config.mlp_output,
name='fnn1'))
self.output_ = tf.layers.dropout(self.output_,
rate=self.config.dropout_rate)
self.logit = tf.layers.dense(inputs=self.output_,
units=2,
name='fnn2')
with tf.variable_scope("optimize-layer"):
self.pred_y = tf.argmax(tf.nn.softmax(self.logit), 1)
# 损失函数，交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
logits=self.logit,
labels=self.y) # 对logits进行softmax操作后，做交叉墒，输出的是一个向量
self.loss = tf.reduce_mean(cross_entropy) # 将交叉熵向量求和，即可得到交叉熵
# 优化器
self.optim = tf.train.AdamOptimizer(
learning_rate=param.BaseConfig.lr).minimize(self.loss)
with tf.name_scope("accuracy"):
# 准确率
correct_pred = tf.equal(
tf.argmax(self.y, 1), self.pred_y
) # 由于input_y也是onehot编码，因此，调用tf.argmax(self.input_y)得到的是1所在的下表
self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
def build_model(self):
with tf.variable_scope("first-CNN-layer"):
self.output_x1_1 = tf.layers.conv1d(
self.input_X1,
filters=self.config.filters_num,
kernel_size=self.config.first_kernel_size,
padding='same',
name='first-cnn1')
self.output_x2_1 = tf.layers.conv1d(
self.input_X2,
filters=self.config.filters_num,
kernel_size=self.config.first_kernel_size,
padding='same',
name='first-cnn2')
with tf.variable_scope("first-interaction"):
interactionModel = mudules.Interaction(3, self.output_x1_1,
self.output_x2_1)
self.inter_x1_1, self.inter_x2_1, self.inter_x2_weight_1 = interactionModel.exeInteraction(
)
with tf.variable_scope("second-CNN-layer"):
self.output_x1_2 = tf.layers.conv1d(
self.inter_x1_1,
filters=self.config.filters_num,
kernel_size=self.config.second_kernel_size,
padding='same',
name='second-cnn1')
self.output_x2_2 = tf.layers.conv1d(
self.inter_x2_1,
filters=self.config.filters_num,
kernel_size=self.config.second_kernel_size,
padding='same',
name='second-cnn2')
with tf.variable_scope("second-interaction"):
interactionModel = mudules.Interaction(3, self.output_x1_2,
self.output_x2_2)
self.inter_x1_2, self.inter_x2_2, self.inter_x2_weight_2 = interactionModel.exeInteraction(
)
with tf.variable_scope("fusion-layer"):
fusionModel = mudules.Fusion(1, self.inter_x2_weight_1,
self.inter_x2_weight_2)
self.fusion_output = fusionModel.exeFusion()
with tf.variable_scope("predict-layer"):
self.output_ = tf.nn.relu(
tf.layers.dense(inputs=self.fusion_output,
units=self.config.mlp_output,
name='fnn1'))
self.output_ = tf.layers.dropout(self.output_,
rate=self.config.dropout_rate)
self.logit = tf.layers.dense(inputs=self.output_,
units=2,
name='fnn2')
with tf.variable_scope("optimize-layer"):
self.pred_y = tf.argmax(tf.nn.softmax(self.logit), 1)
# 损失函数，交叉熵
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
logits=self.logit,
labels=self.y) # 对logits进行softmax操作后，做交叉墒，输出的是一个向量
self.loss = tf.reduce_mean(cross_entropy) # 将交叉熵向量求和，即可得到交叉熵
# 优化器
self.optim = tf.train.AdamOptimizer(
learning_rate=param.BaseConfig.lr).minimize(self.loss)
with tf.name_scope("accuracy"):
# 准确率
correct_pred = tf.equal(
tf.argmax(self.y, 1), self.pred_y
) # 由于input_y也是onehot编码，因此，调用tf.argmax(self.input_y)得到的是1所在的下表
self.acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))