In the 2018 Zoph et al paper, the authors proposed transfereable NAS architecture called NASnet-A (along others) that uses two reduction cells (found by search algorithm on customized search space) in order to transfer the learning from CIFAR-10(32 image size) to Imagenet (using 331/224/299/320 input size). How can we adapt the number of reduction cells used right after the stem convolution for different image sizes?