To find the best network combination, you can refer some literature and look for the existing state of the art network for drowsy detection. Using existing networks like resnet or alexnet is good to start with, but literature will give you more insight to the problem. Check some available network in literature and implement in MATLAB.
For resnet50 or any other pretrained network, you can experiment with varying learning rate, changing optimizer / loss function, change data preprocessing steps and check if the accuracy gets improved. This may or may not be helpful.