About training data and Net structure?

[SyGoing]

The training data can be marked by labelme? , and the training png is index picture? And, If I modified your net to make the output's tensor's shape equal input, there is nothing problem? since it is tow class, why output channel is one instead of two ,like others