Conditional Generative Adversarial Nets
M.Mirza andS.Osindero,arXiv,pp.1-7,nov.2014
Generative adversarial nets(GAN) are extended to conditional GAN. Generator and discriminator are conditioned by some extra information, this can be any kind of auxiliary information including class labels or data. Conditioning is done by feeding the information to both discriminator and generator by adding an extra input layer. In generator , noise and y are combined to produce an image x and in discriminator the x and y is combined.
Generator uses ReLu activation with the final sigmoid unit. The modal is trained using SGD with batch size 100. Initially the learning rate is set to 0.1 and decreased exponentially to 0.000001 and decay factor of 1.00004 and dropout with the probability of 0.5 is applied to both generator and discriminator.
In unconditioned generative models there is no control on modes of the data being generated. In conditioned generative models , it is possible to direct the data generation process.
Generator uses ReLu activation with the final sigmoid unit. The modal is trained using SGD with batch size 100. Initially the learning rate is set to 0.1 and decreased exponentially to 0.000001 and decay factor of 1.00004 and dropout with the probability of 0.5 is applied to both generator and discriminator.
In unconditioned generative models there is no control on modes of the data being generated. In conditioned generative models , it is possible to direct the data generation process.