Advice for training deeper networks on very small datasets

I am trying to build a model with fewer parameters that generalizes well on the flower102 dataset, a very small one in today’s standards. I aim to arrive at an architecture that is smaller, efficient, with fewer parameters, without overfitting. I have read a few papers, such as the EfficientNet, SqueezeNet, and GoogleNet papers. Any advice on how I can approach this research problem?

Maybe pre-training in bigger datasets can do the trick, and add several data augmentation to your goal dataset.

I am currently trying to pretrain the model on the Cifar100 dataset. But then the images in Cifar100 are 32x32 but the flower102 has images of 224x224. Should the goal dataset resolutions also be decreased to 32x32?

I don’t se a problem, you just have to be certain that your model can handle the different image size.