Let us introduce the following setting. Denote by CNN$($in$\_\{$ch$\},$ out$\_\{$ch$\},$ kernel $=($k$\_$h$,$ k$\_$w$),$ stride $=($s$\_$w$,$s$\_$h$),$ padding $=($p$\_$w$,$ ph$))$ the convolutional layer. Here, we use the following notations: in$\_\{$ch$\}-$ the number of input channels, out$\_\{$ch$\}-$ the number of output channels, kernel $-$ dimensions of the convolutional kernel $($k$\_$h $-$ height, k$\_$w $-$ width$),$ stride $-$ convolution pixel step $($two dimensions$),$ padding $-$ number of pixels in two dimensions that are placed to expand the boundary of the input. For further convenience, let us use one value in the case if the two dimensions that are placed to expand the boundary of the input. For further convenience, let us use the one value in the case if the two dimensional parameter takes equal values in each of its dimensions, e$.$g$.$ kernel $=3$ means kernel $=(3,3).$ After convolutional layer we use some nonlinear function, as it is supposed to be$.$ To conclude, CNN$()$ is introduced in the most standardized way, as it seems possible. Also let us consider the standard max polling layer, with kernel size $2$x$2$ and denoted by MLP$.$ Therefore the setting is complete. Consider the following neural network diagram.

CNN$(3,8,3,1,0)->$ MPL $->$ CNN$(8,16,(1,3),1,0)->$ CNN$(16,16,(3,1),1,2)->$ MLP$.$

We put through this model one tensor of form $(3,32,32)($channel$,$ height, width$).$ After that we flatten the output. Thus we get the $1-$d vector. What is the length of this vector?

a$)400$

b$)784$

c$)1024$

d$)1408$