Hao-Wen Dong, Yi-Hsuan Yang

Music and AI Lab,
Research Center for IT Innovation,
Academia Sinica

DANTest

Dicriminative Adversarial Networks (DANs)

Discriminative adversarial networks (DANs) [1] are essentially conditional GANs where both the generator and the discriminator are discriminative models.

system

Based on DANs, we propose a new, simple framework, dubbed DANTest for systematically comparing different adversarial losses.

  1. Build several DANs. For each of them, the generator G takes as input a real sample and outputs a fake label. The discriminator takes as input a real sample with either its true label, or a fake label made by G, and outputs a scalar indicating if the “sample–label” pair is real.
  2. Train the DANs with different component loss functions, regularization approaches or hyperparameters.
  3. Predict the labels of test data by the trained models.
  4. Compare the performance of different models with standard evaluation metrics used in supervised learning.

The DANTest is simple and it is easy to control and extend, which allows us to easily evaluate new adversarial losses. With the DANTest, we are able to conduct a extensive comparative study on different adversarial losses (168 in total) to see how different adversarial losses perform against one another.

Specifically, we consider 10 existing component functions, 2 new component functions propose in light of our theoretical analysis, along with 14 different regularization approaches. Moreover, we use the DANTest to empirically study the effect of the Lipschitz constant, penalty weights momentum terms, and other hyperparameters. Please refer to our paper for the results.