Created by: ARDUJS

环境

• paddle = 1.7.2
• python = 3.7.5

问题

如何构建下面这个lstm，多了T输入

尝试的解决方案

class DecoderCell(layers.RNNCell):
    def __init__(self, hidden_size):
        self.hidden_size = hidden_size
        self.lstm_cell = layers.LSTMCell(hidden_size)

    def call(self, step_input, hidden):
        h, c, t = hidden
        # layers.Print(h)
        i_gate = layers.fc(input=[step_input, t, h], size=self.hidden_size, act='sigmoid')
        f_gate = layers.fc(input=[step_input, t, h], size=self.hidden_size, act='sigmoid')
        z = layers.fc(input=[step_input, t, h], size=self.hidden_size, act='tanh')
        ct = layers.elementwise_mul(f_gate, c) + layers.elementwise_mul(i_gate, z)
        o_gate = layers.fc([step_input, t, h], size=self.hidden_size, act='sigmoid')
        h = layers.elementwise_mul(o_gate, layers.tanh(ct))
        t = layers.fc(h, size=self.hidden_size)
        y = layers.fc(t, size=class_num)
        return y, [h, c, t]

h = layers.fc(pad_feature, size=128, act="tanh")
c = layers.fc(pad_feature, size=128, act="tanh")
t = layers.fc(pad_feature, size=128, act="tanh")
decoder_cell = DecoderCell(128)
decoder_output, _ = layers.rnn(cell=decoder_cell, inputs=pad_feature, initial_states=[h, c, t], time_major=False, sequence_length=seq_len_used)

> 这样导致训练巨慢，可有其他的解决方案，help，大佬！！！