Model Implementation: SEGAN

In this lesson, we are going to implement the discriminator and the generator proposed in the SEGAN paper. While the discriminator takes a clean and a noisy waveform as input, the generator takes in a noisy waveform and a $z$ vector and uses them to produce a denoised version of the noisy waveform.

Discriminator

In SEGAN, the discriminator architecture is similar to DCGAN and comprises a series of convolutions with batch normalization and parametric ReLUs followed by a few dense layers. The main advantage of parametric ReLUs over ReLUs and Leaky ReLUs is that the model and the optimization problem being solved will define the best parameter of the Leaky ReLU.

Let’s start with the necessary imports: