Python实现LSTNet【tensorflow2.0】【时间序列】

糖逗

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[2]
        self.TimeStamp = input_shape[1]
        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):
        #[batchsize, timestramp, countrycount] -> [batchsize, timestramp, CnnChannnel]
        cnn_out = self.CNN(x)
        #[batchsize, timestramp, CnnChannnel] -> [batchsize, GruChannel]
        rnn_out = self.RNN(cnn_out)
        ######################跳跃GRU########################
        #[batchsize, timestramp, CnnChannnel] -> [batchsize, int(-p*IntervalCount), CnnChannnel]
        #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):, :]
        #[batchsize, int(-skip*IntervalCount), CnnChannnel] -> [batchsize, IntervalCount, skip, CnnChannnel]
        cnn_out_stack = tf.reshape(cnn_out_cut, [-1, self.IntervalCount, self.skip, self.CnnChannel])
        #[batchsize, IntervalCount, skip, CnnChannnel] -> [batchsize, skip, IntervalCount, CnnChannnel]
        cnn_out_exchange = tf.transpose(cnn_out_stack, [0, 2, 1, 3])
        #[batchsize, skip, IntervalCount, CnnChannnel] -> [batchsize*skip, IntervalCount, CnnChannnel]
        cnn_out_input = tf.reshape(cnn_out_exchange, [-1, self.IntervalCount, self.CnnChannel])
        #[batchsize*skip, IntervalCount, CnnChannnel]-> [batchsize*skip, GruSkipChannel]
        cnn_skip = self.RNNSkip(cnn_out_input)
        #[batchsize*skip, GruSkipChannel] -> [batchsize, skip*GruSkipChannel]
        cnn_skip_out = tf.reshape(cnn_skip, [-1, self.skip * self.GruSkipChannel])
        #合并RNN和Skip-RNN
        #[batchsize, GruChannel] concate [batchsize, skip*GruSkipChannel] -> [batchsize, GruChannel + skip*GruSkipChannel]
        con_cnn = layers.concatenate([rnn_out, cnn_skip_out], axis = 1)
        #[batchsize, GruChannel + skip*GruSkipChannel] -> [batchsize, CountryDims]
        out = self.Dense(con_cnn)
        #######################线性部分######################
        #线性AR模型
        # highway，模型线性AR
        #[batchsize, timestramp, countrycount] -> [batchsize, hw, countrycount]
        linear = x[:, -self.hw:, :]
        linear = tf.convert_to_tensor(linear)
        #[batchsize, hw, countrycount] -> [batchsize,countrycount, hw]
        linear = tf.transpose(linear, [0, 2, 1])
        #print(linear.shape)
        #[batchsize,countrycount, hw] -> [batchsize*countrycount, hw]
        linear = tf.reshape(linear, [-1, self.hw])
        #[batchsize*countrycount, hw] -> [batchsize*countrycount, 1]
        linear = layers.Dense(1)(linear)
        #[batchsize*countrycount, 1] -> [batchsize, countrycount]
        linear = tf.reshape(linear, (-1, self.CountryDims))
        res = layers.add([out, linear])
        return res
    
model = LSTNet()
model.build((None, 14, 8))
model.summary()       

antonybear

厉害