This repo wraps BigGAN to implement the proposed RCF-GAN. The detailed introduction of the RCF-GAN can be found at here https://github.com/ShengxiLi/rcf_gan.
Follow the steps in the original BigGAN-Pytorch repo: https://github.com/ajbrock/BigGAN-PyTorch
We have provided example scripts. Briefly speaking, the BigRCF-GAN method is enabled by mainly specifying --model RcfGAN, to use the RCF-GAN model, together with --which_train_fn RCFGAN, to enable the corresponding training strategy. Otherwise, you are training your models for unconditional BigGAN generation. The t-net is enabled by default and can be disabled by --t_sigma_num 0. Please also note that the RCF-GAN, in its current form, only supports unconditional generation, so that a self-modulation technique is employed by specifying --unconditional.
- The basis of the BigRCF-GAN is built upon the official BigGAN-Pytorch at https://github.com/ajbrock/BigGAN-PyTorch.
- The implementation of self-modulation is adopted from https://github.com/boschresearch/unetgan.