two questions about RNN

[helloworld1973]

two questions about RNN

[AlexDBlack]

Input modulation, forget, and output gates in LSTM are always sigmoid (hard-coded).
Input and output gates are what you specify there. See the LSTM unit schematic about 60% of the way down this page: http://deeplearning4j.org/lstm

BaseRecurrentLayer is abstract, so you can't instantiate it. Other recurrent layers build on it.

[helloworld1973]

Thank you very much.If I want to use the original RNN which is a feed forward net with time series,not LSTM.Is there anything available to use? @AlexDBlack

[AlexDBlack]

As in a 'vanilla'/ordinary RNN? We haven't got around to implementing that yet. It wouldn't be hard to add (the model itself is quite simple), but performance generally isn't very good compared to LSTM, which is why we haven't implemented it already.

[helloworld1973]

ordinary RNN.I want to change the activation function to make the nets perform better inspired by a paper:A Simple Way to Initialize Recurrent Networks of relu (2015).I intended to do it based on your wonderful work because it includes abundant activation function.