Fig. 4 GAN training paradigm. The discriminator predicted a PIGD score as well as a real/fake label for every walk example input. The generator provided fake walk examples. These outputs were used to compute loss terms for training the discriminator and generator. ( loss_train was also used to train the CNN). See Eqs. (2 –