Question on NoiseInjection

[XavierXiao]

Hi, thanks for the nice implementation. In your implementation for NoiseInjection, you have
self.weight = nn.Parameter(torch.zeros(1, channel, 1, 1))
and you apply the noise by
image + self.weight * noise

But in Facebook's implmentation (https://github.com/facebookresearch/pytorch_GAN_zoo/blob/b75dee40918caabb4fe7ec561522717bf096a8cb/models/networks/styleGAN.py#L35), it looks like they scale the noise once and add the same noise to each channel?

[rosinality]

https://github.com/NVlabs/stylegan/blob/master/training/networks_stylegan.py#L270 Official implementation uses per-channel noise weights.