Separate the linear layer for different features

Your’re right, the parameters in the convolution would be shared… My bad