Identity weights for 2d convolution

I would like to initialise a multi-channel 2d convolution layer such that it simply replicates the input (identity). For a single channel image I know the identity kernel is:

[0, 0, 0
 0, 1, 0
 0, 0, 0]

But how can I do this for a 3 channel image?

Input is [b, 3, h, w] and output is also [b, 3, h, w], so my weight size would be [3, 3, 3, 3].

My best guess which doesn’t work (x should equal y):

x = torch.rand(1, 3, 3, 3)
w = torch.tensor([
	[0., 0., 0.,],
	[0., 1., 0.,],
	[0., 0., 0.,],
]).view(1, 1, 3, 3).repeat(3, 3, 1, 1)
y = nn.functional.conv2d(x, w, bias=None, stride=1, padding=1, dilation=1)
print(torch.allclose(x, y)) # prints False
1 Like

nn.init.dirac_ would do this

2 Likes