Creating RNN in Keras is much easier as compared to the TensorFlow. As you learned in chapter 3, Keras offers both functional and sequential API for creating the recurrent networks. To build the RNN model, you have to add layers from the kera.layers.recurrent module. Keras provides the following kinds of recurrent layers in the keras.layers.recurrent module:
- SimpleRNN
- LSTM
- GRU
Stateful Models
Keras recurrent layers also support RNN models that save state between the batches. You can create a stateful RNN, LSTM, or GRU model by passing stateful parameters as True. For stateful models, the batch size specified for the inputs has to be a fixed value. In stateful models, the hidden state learnt from training a batch is reused for the next batch. If you want to reset the memory at some point during training, it can be done with extra code by calling the model.reset_states...
