Tensorflow, how to access all the middle states of an RNN, not just the last state

Question

My understanding is that tf.nn.dynamic_rnn returns the output of an RNN cell (e.g. LSTM) at each time step as well as the final state. How can I access cell states in all time steps not just the last one? For example, I want to be able to average all the hidden states and then use it in the subsequent layer.

The following is how I define an LSTM cell and then unroll it using tf.nn.dynamic_rnn. But this only gives the last cell state of the LSTM.

import tensorflow as tf
import numpy as np

# [batch-size, sequence-length, dimensions] 
X = np.random.randn(2, 10, 8)
X[1,6:] = 0
X_lengths = [10, 6]

cell = tf.contrib.rnn.LSTMCell(num_units=64, state_is_tuple=True)

outputs, last_state = tf.nn.dynamic_rnn(
    cell=cell,
    dtype=tf.float64,
    sequence_length=X_lengths,
    inputs=X)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())                                 
out, last = sess.run([outputs, last_state], feed_dict=None)

Answer 1

Something like this should work.

import tensorflow as tf
import numpy as np


class CustomRNN(tf.contrib.rnn.LSTMCell):
    def __init__(self, *args, **kwargs):
        kwargs['state_is_tuple'] = False # force the use of a concatenated state.
        returns = super(CustomRNN, self).__init__(*args, **kwargs) # create an lstm cell
        self._output_size = self._state_size # change the output size to the state size
        return returns
    def __call__(self, inputs, state):
        output, next_state = super(CustomRNN, self).__call__(inputs, state)
        return next_state, next_state # return two copies of the state, instead of the output and the state

X = np.random.randn(2, 10, 8)
X[1,6:] = 0
X_lengths = [10, 10]

cell = CustomRNN(num_units=64)

outputs, last_states = tf.nn.dynamic_rnn(
    cell=cell,
    dtype=tf.float64,
    sequence_length=X_lengths,
    inputs=X)

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())                                 
states, last_state = sess.run([outputs, last_states], feed_dict=None)

This uses concatenated states, as I don't know if you can store an arbitrary number of tuple states. The states variable is of shape (batch_size, max_time_size, state_size).