maple728 / da-rnn Goto Github PK

Hi again,

Also wanted to ask about attention layers, in the paper it says
en_w_e = tf.Variable(tf.truncated_normal(shape = [2 * last_hidden_size, config.timestep]), dtype = tf.float32)
en_u_e = tf.Variable(tf.truncated_normal(shape = [config.timestep * config.input_size, config.timestep]), dtype = tf.float32)
en_v_e = tf.Variable(tf.truncated_normal(shape = [config.timestep, 1]))

Are parameters learned through a MLP model, but I dont see any dense layer in your code, you just initialize the weights, can you help me understand how these weights are adjusted through training?

Thanks

计算最终预测值时，论文中给出的方法是对解码器的隐状态和context_vector进行concatenation。您的代码中是X_encoders，是我哪里理解错了吗？

Hi Great work thank you.

I wanted to ask you the reason why you including last 10 points of train set to valid set also last 10 points of valid set to test set? Is there a specific reason why are you doing like that?

I am referring to here:
return dataset[ :, : train_end_index], dataset[:, train_end_index - config.timestep : valid_end_index], dataset[:, valid_end_index - config.timestep : -1]

Thanks a lot