Python实现LSTNet【tensorflow2.0】【时间序列】
论文链接：https://arxiv.org/pdf/1703.07015.pdf参考链接：https://zhuanlan.zhihu.com/p/61795416
class LSTNet(keras.Model):
def __init__(self, ):
super(LSTNet, self).__init__()
self.CnnChannel = 32#CNN输出的channel数
self.CnnKernelSize = 5#CNN中Kernel的大小
self.GruChannel = 16#GRU输出的channel数
self.GruSkipChannel = 16#GRU_Skip的输出channel数
self.skip = 7#时间跳跃的跨度
self.hw = 7#AR线性窗口
def build(self, input_shape):
self.CountryDims = input_shape
self.TimeStamp = input_shape
self.IntervalCount = int(self.TimeStamp / self.skip)
#################非线性层###################
self.CNN = layers.Conv1D(filters = self.CnnChannel,
kernel_size = self.CnnKernelSize,
activation = "relu", #dropout = 0.5,
padding = "same")
#非跳跃的RNN层
self.RNN = layers.GRU(units = self.GruChannel,
dropout = 0.5, unroll = True)
#跳跃的RNN层
self.RNNSkip = layers.GRU(units = self.GruSkipChannel,
dropout = 0.5, unroll = True)
self.Dense = layers.Dense(units = self.CountryDims)
#################线性层######################
super(LSTNet, self).build(input_shape)
def call(self, x, training = None):
# ->
cnn_out = self.CNN(x)
# ->
rnn_out = self.RNN(cnn_out)
######################跳跃GRU########################
# ->
#给定周期跨度下，可能存在原时间戳长度不能整除的情况，所以这里用int(-p*IntervalCount)
cnn_out_cut = cnn_out[:, int(-self.skip * self.IntervalCount):, :]
# ->
cnn_out_stack = tf.reshape(cnn_out_cut, [-1, self.IntervalCount, self.skip, self.CnnChannel])
# ->
cnn_out_exchange = tf.transpose(cnn_out_stack, )
# ->
cnn_out_input = tf.reshape(cnn_out_exchange, [-1, self.IntervalCount, self.CnnChannel])
#->
cnn_skip = self.RNNSkip(cnn_out_input)
# ->
cnn_skip_out = tf.reshape(cnn_skip, [-1, self.skip * self.GruSkipChannel])
#合并RNN和Skip-RNN
# concate ->
con_cnn = layers.concatenate(, axis = 1)
# ->
out = self.Dense(con_cnn)
#######################线性部分######################
#线性AR模型
# highway，模型线性AR
# ->
linear = x[:, -self.hw:, :]
linear = tf.convert_to_tensor(linear)
# ->
linear = tf.transpose(linear, )
#print(linear.shape)
# ->
linear = tf.reshape(linear, [-1, self.hw])
# ->
linear = layers.Dense(1)(linear)
# ->
linear = tf.reshape(linear, (-1, self.CountryDims))
res = layers.add()
return res
model = LSTNet()
model.build((None, 14, 8))
model.summary() 厉害
页:
[1]