When running a batch of size $32$ through an L$-$layer deep neural network, where each sample could possibly belong to one of $3$ output categories, the shape of the raw scores matrix is $\_\_\_\_\_\_\_\_.$