Question about implementing trainable PCEN layer using TensorFlow

[lydiaji]

Thanks for sharing PCEN implementation using Pytorch. Currently I'm trying to use TensorFlow to add trainable PCEN layer before KWS model, but the result is not so good. Would it be a problem that all training data is 1 second and then fed to the net? And I'm not sure about the initialization of parameters are correct. I just use the initialization defined in the paper. But I'm not clear how to first use log values then take exponentials. Could you please give me some guide? Thanks so much!