Created by: pkuyym

Current implementation has much space to improve:

1. Use process instead of thread for batch assembling task.
2. Use process instead of thread for ordered data feeding task.
3. In batch iterator, should avoid to sharing raw numpy array between processes.