yolo_prune

1. set weights < 0.01 to 0 (set PRUNE = 1 in makefile);

2. pruning output feature maps according to BN scales (set SCALE_L1 = 1 in makefile)

3. conv connection sparsification by using a mask of 0 and 1 (set MASK = 1 in makefile)

4. parallel computation (set MULTI_CORE = 1 in makefile)

The first function is based on the work of https://github.com/hjimce/compress_yolo

The second function is pruning feature maps according to BN scales parameter,

and the regularization method is L1 NORM

test result is:
YOLO2
Before Runtime: 15s, mAP: 0.62
After Runtime: 2s, mAP: 0.57
when I test tiny-yolo, the runtime can reduced to < 1s in CPU

If you want to use my code, please let me know