Hello, I'm trying to implement a CNN from scratch and I'd like to understand how m= any 2d matrices are being produced during a single stage of the convolution= operation. If image components (RGB) are split into three 2d matrices (W x H x Compone= nt), and a convolution operation is applied to each of those (over a 3 x 3 = 2d kernel, for example), does that mean that for each kernel I would get three 2d output matrices? If the above is true, and each convolution layer can have multiple kernels = (let's say 8 kernels per layer), does that mean that the first layer would = produce 8 * 3 =3D 24 matrices, and each subsequent layer would produce 8 ti= mes more matrices (2nd layer produces 8 * 24 =3D 192 matrices, 3rd produces= 8 * 192 =3D 1536, etc.) ? I would appreciate it if someone could clarify this for me. Thanks a lot, M

0 |

12/18/2016 7:37:20 PM