if __name__ == "__main__":
D_input = 39
D_label = 24
learning_rate = 7e-5
num_units = 1024
EPOCH = 100
train_data = ""
test_data = " "
# 样本的输入和标签
inputs = tf.placeholder(tf.float32, [None, None, D_input], name="inputs")
labels = tf.placeholder(tf.float32, [None, D_label], name="labels")
# 实例LSTM类
rnn_cell = LSTM(inputs, D_input, num_units, orthogonal_initializer)
# 调用scan计算所有hidden states
rnn0 = rnn_cell.all_steps()
# 将3维tensor [n_steps, n_samples, D_cell]转成 矩阵[n_steps*n_samples, D_cell]
# 用于计算outputs
rnn = tf.reshape(rnn0, [-1, num_units])
# 输出层的学习参数
W = weight_init([num_units, D_label])
b = bias_init([D_label])
output = tf.matmul(rnn, W) + b
# 损失
loss = tf.reduce_mean((output-labels)**2)
# 训练所需
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)
# 建立session并实际初始化所有参数
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
train_epoch(EPOCH, train_data, test_data)
