Training own image datasets using caffenet

[Sandesh Sharma]

I want to train my own image datasets using caffenet(caffe imagenet example). Caffenet uses 256*256 images. However I want to train my datasets on different dimensions. For same image dataset, I want to train on 28*28, 64*64, 128*128 and 256*256 image datasets and check the difference in accuracy. For this, do I have to change the train_val.prototxt? Do I have to change values on different layers? I am using a binary classifier, so output of last layer should be 2 right? I am confused on these things. Can anyone provide some steps to train own datasets for different image size?

P.S. I have two directories A and B. Each directory contain 20,000 images. Training, validation and testing datasets will be splitted from these two directories.