An improvement on regression and data reading-revision

[tujie-jiangye]

The improvement of this version is mainly aimed at multiple label regression problems. By improving LMDB reading and writing, the Intel caffe LMDB database can directly support multi label regression without the h5py database as an intermediary. Through this improvement, we find that some native Caffe regression loss is not convergent or slow convergence, this improvement has good adaptability, and shows excellent results on similar problems.