def cosine_error(x): #x=[x1,x2,x3,x4] ,xi.shape=(batch_size,input_dim)
cos1=cosine(x[0],x[1]) #cos shape=(batch_size,)
cos2=cosine(x[0],x[2])
cos3=cosine(x[0],x[3])
cos4=cosine(x[0],x[4])
cos5=cosine(x[0],x[5])
cos6=cosine(x[0],x[6])
delta=5
p=K.exp(cos1*delta)/(K.exp(cos1*delta)+K.exp(cos2*delta)+K.exp(cos3*delta)+K.exp(cos4*delta)+K.exp(cos5*delta)+K.exp(cos6*delta)) #softmax
f=-K.log(p) #objective function：-log #f.shape=(batch_size,)
return K.reshape(f,(K.shape(p)[0],1)) #return.sahpe=(batch_size,1)
def call(self,inputs,mask=None):
#w_c=K.repeat(self.W_c,self.input_num)
#w_m=K.repeat(self.W_m,self.input_num)
x=inputs[0]
mem_vector=inputs[1]
c=K.dot(x,self.W_c)+self.b_c #context向量
m=K.dot(x,self.W_m)+self.b_m #memory向量
mem_vec=K.repeat(mem_vector,self.input_num) #与问题进行内积
m=K.sum(m*mem_vec,axis=2,keepdims=False)
s=K.softmax(m) #softmax
s=K.reshape(s,(-1,self.input_num,1))
ctx=self.activation(c*s)
return ctx#self.activation(ctx)
def call(self, inputs, mask=None):
#w_c=K.repeat(self.W_c,self.input_num)
#w_m=K.repeat(self.W_m,self.input_num)
x = inputs[0]
mem_vector = inputs[1]
c = K.dot(x, self.W_c) + self.b_c #context向量
m = K.dot(x, self.W_m) + self.b_m #memory向量
mem_vec = K.repeat(mem_vector, self.input_num) #与问题进行内积
m = K.sum(m * mem_vec, axis=2, keepdims=False)
s = K.softmax(m) #softmax
s = K.reshape(s, (-1, self.input_num, 1))
ctx = self.activation(c * s)
return ctx #self.activation(ctx)
def LSTNet(trainX1, trainX2, trainY, config):
input1 = Input(shape=(trainX1.shape[1], trainX1.shape[2]))
conv1 = Conv1D(filters=48, kernel_size=6, strides=1,
activation='relu') # for input1
# It's a probelm that I can't find any way to use the same Conv1D layer to train the two inputs,
conv2 = Conv1D(filters=48, kernel_size=6, strides=1,
activation='relu') # for input2
conv2.set_weights(conv1.get_weights()) # at least use same weight
conv1out = conv1(input1)
lstm1out = CuDNNLSTM(64)(conv1out)
lstm1out = Dropout(config.dropout)(lstm1out)
input2 = Input(shape=(trainX2.shape[1], trainX2.shape[2]))
conv2out = conv2(input2)
lstm2out = CuDNNLSTM(64)(conv2out)
lstm2out = Dropout(config.dropout)(lstm2out)
lstm_out = concatenate([lstm1out, lstm2out])
output = Dense(trainY.shape[1])(lstm_out)
#highway 使用Dense模拟AR自回归过程，为预测添加线性成份，同时使输出可以响应输入的尺度变化。
highway_window = config.highway_window
#截取近3个窗口的时间维 保留了所有的输入维度
z = Lambda(lambda k: k[:, -highway_window:, :])(input1)
z = Lambda(lambda k: K.permute_dimensions(k, (0, 2, 1)))(z)
z = Lambda(lambda k: K.reshape(k,
(-1, highway_window * trainX1.shape[2])))(z)
z = Dense(trainY.shape[1])(z)
output = add([output, z])
output = Activation('sigmoid')(output)
model = Model(inputs=[input1, input2], outputs=output)
return model