TensorFlow dynamic_rnn state

Question

My question is about the TensorFlow method tf.nn.dynamic_rnn. It returns the output of every time step and the final state.

I would like to know if the

Answer 1

tf.nn.dynamic_rnn returns two tensors: outputs and states.

The outputs holds the outputs of all cells for all sequences in a batch. So if a particular sequence is shorter and padded with zeros, the outputs for the last cells will be zero.

The states holds the last cell state, or equivalently the last non-zero output per sequence (if you're using BasicRNNCell).

Here's an example:

import numpy as np
import tensorflow as tf

n_steps = 2
n_inputs = 3
n_neurons = 5

X = tf.placeholder(dtype=tf.float32, shape=[None, n_steps, n_inputs])
seq_length = tf.placeholder(tf.int32, [None])

basic_cell = tf.nn.rnn_cell.BasicRNNCell(num_units=n_neurons)
outputs, states = tf.nn.dynamic_rnn(basic_cell, X, sequence_length=seq_length, dtype=tf.float32)

X_batch = np.array([
  # t = 0      t = 1
  [[0, 1, 2], [9, 8, 7]], # instance 0
  [[3, 4, 5], [0, 0, 0]], # instance 1
])
seq_length_batch = np.array([2, 1])

with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())
  outputs_val, states_val = sess.run([outputs, states], 
                                     feed_dict={X: X_batch, seq_length: seq_length_batch})

  print('outputs:')
  print(outputs_val)
  print('\nstates:')
  print(states_val)

This prints something like:

outputs:
[[[-0.85381496 -0.19517037  0.36011398 -0.18617202  0.39162001]
  [-0.99998015 -0.99461144 -0.82241321  0.93778896  0.90737367]]

 [[-0.99849552 -0.88643843  0.20635395  0.157896    0.76042926]
  [ 0.          0.          0.          0.          0.        ]]]  # because len=1

states:
[[-0.99998015 -0.99461144 -0.82241321  0.93778896  0.90737367]
 [-0.99849552 -0.88643843  0.20635395  0.157896    0.76042926]]

Note that the states holds the same vectors as in output, and they are the last non-zero outputs per batch instance.