In an LSTM network once the training is done we have the final states available. I have two questions on further usage of these LSTM tuple states ( c and h). (i) How do we use these states for testing the model against validation datasets ? (ii) If we need to predict the class then there are two suggested approaches in this forum. Not sure which one is correct.
a. Use the last state values (from last batch observation) and use it for prediction. Tensorflow save final state of LSTM in dynamic_rnn for prediction
b. Reset the states to zero and do the prediction. Tensorflow, best way to save state in RNNs?
