Decoder Initial State and Input

[shubhamagarwal003]

Hi @RobRomijnders ,
Thanks for sharing your code. I am very new to RNNs so forgive me for a silly question. In your decoder I am a little confused about the inputs and initial state. Shouldn't the input be the one obtained from the encoder (the latent space)? It seems in your code

initial_state_dec = tuple([(z_state, z_state)] * num_layers)
dec_inputs = [tf.zeros([batch_size, 1])] * sl 
outputs_dec, _ = tf.contrib.rnn.static_rnn(cell_dec, inputs=dec_inputs,
                                                       initial_state=initial_state_dec)

the initial state is obtained from the encoder and not the inputs? So if you could please explain me the reason for that?

[tejaslodaya]

Hi @shubhamagarwal003 ,
I know its really late for the answer. But if anyone comes here and is stumbled upon the same question, it may be of some help.

In the section Lat_2_dec, latent_vector (z_mu) is passed through a linear layer to obtain z_state.
Now, z_state is of hidden_size dimension, which serves as an initial hidden layer in the decoder.
That's what initial_state_dec is meant by.

And for the inputs for the decoder, they're initialized by zeros, given by dec_inputs.
So, we have inputs and hidden of a decoder.

Let me know if you have any questions.