For image-net,reduce image size to 32x32 and set number of layers to 12 and number of filters to 32

I have tried to search architectures on cifar100, but validation accuracy was so low : 0.59~0.64

Also the found architecture have only skip-connection and pooling layers.

Do you have any recommendation for tuning parameters for searching on cifar100 / imagenet?

I recommend you to check Progressive DARTS, where they raised a clever way to limit the skip-connections. Their work is based on DARTS so it won't be very hard to get hands on.

GL,