Change first convolution layer input channels. In case: in_channels == 1 or in_channels == 2 -> reuse original weights in_channels > 3 -> make random kaiming normal initialization
(model, in_channels)
Source from the content-addressed store, hash-verified