Open raindrop313 opened 1 year ago
Hi raindrop,
Because large λ will compromise the performance for input without perturbations, we adjust the weight to make the final clean performance around 28.0(±0.2) for Rain100H and 28.5(±0.2) for RainCityscape in terms of PSNR.
For your reference, we set it around 0.15 for RainCityscape and Rain100H in the experiment for our proposed model.
Hi, how did you set the λ in the adversarial loss during training