The idea behind noisy CycleGAN is to train the model to generate reconstruction results robust to high-frequency perturbations which lie in the same frequency spectre with the self-adversarial structured noise. After the input image has been mapped to the target domain, the Gaussian noise is added to the translation result before reconstruction of the input. The noisy CycleGAN model is stored in the file cycle_gan_noisy_model.py.
In addition to the parameters of the original CycleGAN, noisy CycleGAN an additional parameter noise_std that sets the standard deviation of a zero-mean Gaussian nose added to the translated image before reconstruciton.
The command for training the model is:
python train.py --model cycle_gan_noisy --dataroot /path/to/data --noise_std *some_double less than 1*