CuDNNLSTM

[GameLogist]

I believe CuDNNLSTM will be beneficial for the model.

[TheErk]

I tested this on my side and CuDNNLSTM seems to work as a drop-in replacement for LSTM. This only works for keras tensorflow backend though.

That said in my current test case using CuDNNLSTM instead of LSTM make learning phase go 3 times faster. I don't took time to thoroughly cross check results between the 2 variants after learning.

Perhaps one could programmatically check backend and use CuDNNLSTM whenever

keras.backend.backend()=='tensorflow'

@GameLogist thank you for your suggestion I'll use it when possible.