I’ve noted that the critic’s (discriminator) loss is always negative, when training the model from notebook lesson7-wgan.ipynb . Why does hat happen? Does critic’s loss, in this case, refers to an average error, without using squared or absolute values? Or, maybe, should I simply consider that the training process aims at minimizing the absolute value of disc_loss (i.e. consider that the “real” loss for the critic would be the value printed without the minus sign) ?
1 Like