Open YueXin18 opened 10 months ago
With self supervision, the context inpainting model will try to create a mask that is more difficult than what the model can reconstruct, increasing the loss. This will result in negative loss to maximize the loss from the context inpainting model such that the model can generate better reconstruction, just like an adversarial training.
May I ask what is the reason for a negative loss value when training a single InfNet model with self supervision?