Method for training a generative adversarial network (GAN), generative adversarial network, computer program, machine-readable memory medium, and device
A method for training a generative adversarial network, in particular a Wasserstein generative adversarial network. The generative adversarial network includes a generator and a discriminator, the generator and the discriminator being artificial neuronal networks. The method includes training the discriminator. In the step of training the discriminator, a parameter of the discriminator is adapted as a function of a loss function, the loss function including a term that represents the violation of the Lipschitz condition as a function of a first input datum and a second input datum and as a function of a first output of the discriminator when processing the first input datum and a second output of the discriminator when processing the second input datum, the second input datum being created starting from the first input datum by applying the method of the virtual adversarial training.
1. A method for training a Wasserstein generative adversarial network, the generative adversarial network including a generator and a discriminator, the generator and the discriminator being artificial neuronal networks, the method including the following:
training the discriminator, the training of the discriminator including adapting a parameter of the discriminator as a function of a loss function, the loss function including a term that represents a violation of a Lipschitz condition as a function of a first input datum and a second input datum, and as a function of: (i) a first output of the discriminator when processing the first input datum, and (ii) a second output of the discriminator when processing the second input datum, the second input datum being created starting from the first input datum by applying a method of a virtual adversarial training.
2. The method as recited in claim 1 , wherein the first input datum is either retrieved from a data memory for real training data or generated using the generator.
3. The method as recited in claim 1 , wherein the first input datum is changed into its adversarial direction for creation while applying the method of the virtual adversarial training, the adversarial direction being approximated by applying a power iteration.
4. The method as recited in claim 1 , wherein the method includes a first step of training the generator and a second step of training the generator, multiple iterations of the step of the training of the discriminator being carried out between the first step of training the generator and the second step of training the generator.
5. The method as recited in claim 1 , wherein the discriminator is near 1-Lipschitz and near optimal.
6. A generative adversarial network, comprising:
a generator; and
a discriminator;
wherein the generator and the discriminator are artificial neuronal networks, the discriminator being trained by adapting a parameter of the discriminator as a function of a loss function, the loss function including a term that represents a violation of a Lipschitz condition as a function of a first input datum and a second input datum, and as a function of: (i) a first output of the discriminator when processing the first input datum, and (ii) a second output of the discriminator when processing the second input datum, the second input datum being created starting from the first input datum by applying a method of a virtual adversarial training.
7. The generative adversarial network as recited in claim 6 ,
wherein the discriminator is near 1-Lipschitz and near optimal.
8. A non-transitory machine-readable memory medium on which is stored a computer program for training an artificial neuronal network including a generator and a discriminator, the computer program, when executed by a computer, causing the computer to perform:
training the discriminator, the training of the discriminator including adapting a parameter of the discriminator as a function of a loss function, the loss function including a term that represents a violation of a Lipschitz condition as a function of a first input datum and a second input datum, and as a function of: (i) a first output of the discriminator when processing the first input datum, and (ii) a second output of the discriminator when processing the second input datum, the second input datum being created starting from the first input datum by applying a method of a virtual adversarial training.
9. A device configured to train a Wasserstein generative adversarial network, the generative adversarial network including a generator and a discriminator, the generator and the discriminator being artificial neuronal networks, the device configured to:
train the discriminator, the training of the discriminator including adapting a parameter of the discriminator as a function of a loss function, the loss function including a term that represents a violation of a Lipschitz condition as a function of a first input datum and a second input datum, and as a function of: (i) a first output of the discriminator when processing the first input datum, and (ii) a second output of the discriminator when processing the second input datum, the second input datum being created starting from the first input datum by applying a method of a virtual adversarial training.