After instantiating a 2D convolution with
conv = nn.Conv2d(8, 8, 3, bias=False), whose member
bias should be
None, is it able to give
conv a legal bias again (whether with random initialization or determined values)?
I observed that
bias in other default convolution modules is of the type
Parameter, so I suspect there are extra procedures beyond simply
conv.bias = torch.tensor(...).