qixianbiao

You should first detect and align faces for both CASIA and LFW. You can train the model and evaluate the performance on LFW each 1000 steps like Facenet does https://github.com/davidsandberg/facenet...

Because, the shape of human face is visually like a rectangle instead a square. Why 112*96 becuase 112/16=7, and 96/16=6, it is reasonable for the network structure.

As you mentioned, you have trained the model, how along does it take for training 50 epoches?? Can you give me some details? Thanks in advance.