I complete the SN by Keras. welcome to advice

[Bingohong]

location in SN by Keras

[pluo911]

Thanks for your interest. SN benefits from adding 0.5 dropout in the last layer of hidden features, but GN and BN might not. The improvement depends on the generation error (val_err - train_err) of your problem. The training time of SN can be improved by using CUDA. We may release this implementation in the future.

[Bingohong]

thanks to apply the dropout trick. i expect the CUDA accelerated version.