In TensorFlow 0.11, the default behavior is for model states to be represented as tuples; forcing the old concatenated behavior using state_is_tuple=False prints a deprecation warning.
Adapting the code to work with state_is_tuple=True turned out to be easier than I expected, because the TensorFlow example script ptb_word_lm.py is structured very similarly to char-rnn-tensorflow and was refactored back in August to use tuple states. This pull request ports the changes from that example to char-rnn-tensorflow.
In TensorFlow 0.11, the default behavior is for model states to be represented as tuples; forcing the old concatenated behavior using
state_is_tuple=False
prints a deprecation warning.Adapting the code to work with
state_is_tuple=True
turned out to be easier than I expected, because the TensorFlow example scriptptb_word_lm.py
is structured very similarly to char-rnn-tensorflow and was refactored back in August to use tuple states. This pull request ports the changes from that example to char-rnn-tensorflow.