Contents
What is a Wasserstein GAN?
The Wasserstein Generative Adversarial Network, or Wasserstein GAN, is an extension to the generative adversarial network that both improves the stability when training the model and provides a loss function that correlates with the quality of generated images.
What is critic in GAN?
Generative Adversarial Network (GAN) A discriminator D estimates the probability of a given sample coming from the real dataset. It works as a critic and is optimized to tell the fake samples from the real ones.
What is vanilla GAN?
Vanilla GAN: This is the simplest type GAN. Here, the Generator and the Discriminator are simple multi-layer perceptrons. In vanilla GAN, the algorithm is really simple, it tries to optimize the mathematical equation using stochastic gradient descent.
How is Wasserstein distance used to train Gan?
This conceptual shift is motivated mathematically using the earth mover distance, or Wasserstein distance, to train the GAN that measures the distance between the data distribution observed in the training dataset and the distribution observed in the generated examples.
How does the Wasserstein loss help the critic?
The Wasserstein loss encourages the critic to separate these numbers. We can also reverse the situation and encourage the critic to output a large score for real images and a small score for fake images and achieve the same result. Some implementations make this change.
How to implement Wasserstein loss for generative data?
The loss function can be implemented by multiplying the expected label for each sample by the predicted score (element wise), then calculating the mean. The above function is the elegant way to implement the loss function; an alternative, less-elegant implementation that might be more intuitive is as follows:
Is the Wasserstein loss function implemented in keras?
In the Keras deep learning library (and some others), we cannot implement the Wasserstein loss function directly as described in the paper and as implemented in PyTorch and TensorFlow. Instead, we can achieve the same effect without having the calculation of the loss for the critic dependent upon the loss calculated for real and fake images.