I'm reading a paper using GAN to process illumination in image. In the paper, the author mentioned using the SSIM loss for quality evaluation of the reconstructed image. However, in the SSIM loss, there is a hyperparam C1,C2, which depend on the dynamic value of the pixels (the author said it is 1 for their model). However, the paper doesn't mention the output layer of the Generator, which i assume that it is the tanh function. But tanh's output value would be between -1 and 1 (dynamic value = 2), which contradict with the SSIM hyperparam i have mentioned. Can anyone explain this for me?
GAN generators can have any type of output activation function. Some have shown to work better than others, such as TanH, but sigmoid is also often used.
The key here is that it depends on your real data. If your real data is in the feature space $[0, 1]^n$, then your output activation should probably be a sigmoid function. If your data resides in a $[-1, 1]^n$ space, it should be a TanH function. However, if your data lies in $[-\infty, \infty]^n$, you could opt even for a linear output layer. The generator simply should be able to generate in the same space as the real data that you throw towards the discriminator.