Tensorflow: Modify Calculation of Hidden State

Question

I would like to slightly modify the hidden state calculated (and consequently re-inputed) at each time step of an LSTM. For example, something as simple as adding a constant vector to the hidden state that is produced by the regular LSTM process. I don't intend to modify the regular gate mechanics -- I view this more as a calculation of the final hidden state that is normally created by an LSTM. I feel creating my own LSTM cell is overdoing it. Alternatively, the OutputProjectionWrapper seems like it could achieve this goal, though I'm not intending to create a new output, just modify the hidden state.

Answer 1

You can modify the state of the LSTM between time steps if you are using a for loop to process your sequences, since an RNN cell in TensorFlow takes two arguments: the new input and the hidden state:

lstm = rnn_cell.BasicLSTMCell(lstm_size)

initial_state = state = tf.zeros([batch_size, lstm.state_size])

with tf.variable_scope("RNN"):
  for time_step in range(num_steps):
    if time_step > 0: tf.get_variable_scope().reuse_variables()

    # modify the state
    modified_state = state + 1

    output, state = lstm(words[:, i], modified_state)


final_state = state