update from old[repo]different baseline, mainly including local feature alignment, attention and multi-scale *data consists of three populardataset,such as Market1501,Duke and cuhk
##instructions
- based on PCB method, achieving local feature alignment
- based on attetion method, trying to aligning trainable weights into multi-scale network