How to understand the MeanShift layer?

[Alxemade]

Hi, @thstkdgus35
Thanks for sharing your code. I read your code, but I feel confused about the MeanShift layerEDSR-PyTorch/src/model/common.py. The MeanShift layer defined as follows:

class MeanShift(nn.Conv2d):
    def __init__(
        self, rgb_range,
        rgb_mean=(0.4488, 0.4371, 0.4040), rgb_std=(1.0, 1.0, 1.0), sign=-1):

        super(MeanShift, self).__init__(3, 3, kernel_size=1)
        std = torch.Tensor(rgb_std)
        self.weight.data = torch.eye(3).view(3, 3, 1, 1) / std.view(3, 1, 1, 1)
        self.bias.data = sign * rgb_range * torch.Tensor(rgb_mean) / std
        self.requires_grad = False

In the edsy.py, you set sign=-1 in the first MeanShift layer to sub the mean, and you set sign=1 in the last layer to add the mean. Is that correct? Can you share me the details this layer such as the meaning of self.weight.data and self.bias.data.
Thank you!

[sanghyun-son]

Hello.

The meanshift layer is a module that adds/subtracts channel-wise mean from the input/output images.

It is equivalent to:

x[:, 0, :, :] -= red_channel_mean
x[:, 1, :, :] -= green_channel_mean
x[:, 2, :, :] -= blue_channel_mean

Since I do not prefer such notations, I inherited the convolutional layer to implement it.

If you check the code carefully, you will notice that only bias matters!

Thank you.

[Alxemade]

That's cool! Thank you for your explanation.

[ninesun127]

I have also meet this problem when training on my own dataset,what is the meaning of the std in code?
https://github.com/yulunzhang/RCAN/issues/60