A GAN variant that uses a Wasserstein-based objective to improve training behavior and stability.
Wasserstein GAN (WGAN) improves GAN training by reducing mode collapse and vanishing-gradient problems. It changes the discriminator into a single-value scorer without a sigmoid output. section 9.0section 9.3
The core implementation choices are:
Use a discriminator loss based on a Wasserstein objective rather than the original GAN formulation. section 9.3
Keep the discriminator output to one value, but remove its sigmoid activation to reduce vanishing-gradient issues. section 9.3
Train the generator with lossG = −D(G(z)); it only needs the discriminator’s score for generated samples. section 9.3.2
Use WGAN-GP when a more consistent training approach is preferred; the added discriminator penalty helps prevent it from becoming too effective. section 9.9
How Inside Deep Learning: Math, Algorithms, Models covers it
How deep it goes
Defines discriminator and generator losses and removes the sigmoid, but defers the mathematics deriving the Wasserstein improvement.
Examples
Uses the discriminator score D(G(z)) and generator loss −D(G(z)) as the central implementation example.
What it assumes
Assumes familiarity with GAN generators, discriminators, adversarial training, and mode collapse.
What you can do after
Apply WGAN losses and discriminator-output changes while understanding their role in reducing training failures.
How it connects
Connects GANs with mode collapse, vanishing gradients, discriminators, generators, and WGAN-GP.
Learn Generative AI with PyTorchA GAN variant that uses the Wasserstein distance as its training objective to improve stability and provide a more informative critic score.
Wasserstein distanceuses · Wasserstein distance is a distance between probability distributions defined by the minimum cost of transporting probability mass between them under a specified ground distance.
generative adversarial networkalternative to · A generative adversarial network is a generative model with a generator and discriminator trained in competition, enabling the generator to produce samples resembling a target dataset or distribution.
mode collapseaddresses · Mode collapse is a generative-model failure in which outputs concentrate on a narrow range of patterns and fail to represent the diversity of the target distribution.
discriminatorcontrasts with · A discriminator is the model component of a generative adversarial network that distinguishes real data from data generated by its generator.
regularizationuses · Regularization is a modeling technique that constrains or penalizes solutions during fitting or optimization to discourage undesirable complexity or behavior and improve generalization or estimation.
gradient penaltyused by · A gradient penalty is a regularization loss term that penalizes undesirable gradients during model optimization.
This explainer was generated by Manning's knowledge graph from the text of Inside Deep Learning: Math, Algorithms, Models. The section links point to the parts of the book it draws on.
