a question about the sigmoid function in the affine coupling layer

[ShipengFu]

Hi, thank you for your amazing work. As you mentioned, using the sigmoid function in the affine coupling layer is benefical to the training. I was wondering why you shifted the sigmoid s = F.sigmoid(log_s + 2). Thanks a lot.

[rosinality]

Shifting will make sigmoid values towards 1, so it is helpful for gradient flow at initializations.

[ShipengFu]

Shifting will make sigmoid values towards 1, so it is helpful for gradient flow at initializations.

Thank you so much for your reply~