Why does the activation should product sqrt(2)?

[joe1chief]

StyleGAN2_PyTorch/network/stylegan2.py

Line 95 in 37741fb

x = x * np.sqrt(2) # original repo def_gain=np.sqrt(2).

[tomguluson92]

Hi, This way is same with original TF version， used in Leaky ReLU.

[joe1chief]

Hi, This way is same with original TF version， used in Leaky ReLU.
Why did you comment on the following code? Is the tanh() necessary?

StyleGAN2_PyTorch/network/stylegan2.py

Line 913 in 37741fb

# [-1, 1]

[tomguluson92]

Hi, This way is same with original TF version， used in Leaky ReLU.
Why did you comment on the following code? Is the tanh() necessary?

StyleGAN2_PyTorch/network/stylegan2.py

Line 913 in 37741fb

# [-1, 1]

I try to restrict the output data fall in [-1, 1], which the original TF version didn't do that.