modify rnn_train.py script, vad_gru and denoise_gru will use relu as …

[pyu1538]

…activation, so it matches the paper

[tracek]

Doing so will prevent the kernel from using cuDNN version of GRU, defaulting to the generic implementation. That in turn will severely slow down the training (assuming you're using GPU for the training).

[pyu1538]

Doing so will prevent the kernel from using cuDNN version of GRU, defaulting to the generic implementation. That in turn will severely slow down the training (assuming you're using GPU for the training).

Got it. I trained it without GPU, so I didn't feel slow down. RNNoise demo is also trained by "relu", it's better to use exact same model for comparison.