se convolution (instead of a linear layer). This avoids a transpose operation (which may be
        expensive on resource-constrained devices) that may be required to convert the unfolded tensor from
        channel-first to channel-last format in case of a linear layer.
    rd