Efficient Conditional GAN Transfer with Knowledge Propagation across Classes

Generative adversarial networks (GANs) have shown impressive results in both\nunconditional and conditional image generation. In recent literature, it is\nshown that pre-trained GANs, on a different dataset, can be transferred to\nimprove the image generation from a small target data. The same, however, has\nnot been well-studied in the case of conditional GANs (cGANs), which provides\nnew opportunities for knowledge transfer compared to unconditional setup. In\nparticular, the new classes may borrow knowledge from the related old classes,\nor share knowledge among themselves to improve the training. This motivates us\nto study the problem of efficient conditional GAN transfer with knowledge\npropagation across classes. To address this problem, we introduce a new GAN\ntransfer method to explicitly propagate the knowledge from the old classes to\nthe new classes. The key idea is to enforce the popularly used conditional\nbatch normalization (BN) to learn the class-specific information of the new\nclasses from that of the old classes, with implicit knowledge sharing among the\nnew ones. This allows for an efficient knowledge propagation from the old\nclasses to the new ones, with the BN parameters increasing linearly with the\nnumber of new classes. The extensive evaluation demonstrates the clear\nsuperiority of the proposed method over state-of-the-art competitors for\nefficient conditional GAN transfer tasks. The code is available at:\nhttps://github.com/mshahbazi72/cGANTransfer\n