The offsets layer has num_group*kernel_width*kernel_height*2 channels, but how the 4D offsets flattened to 1D channel?