hjimce / compress_yolo

对yolo模型进行网络压缩、修剪等实验

compress_yolo 对yolo模型进行网络压缩、修剪等实验

1.git clone https://github.com/hjimce/compress_yolo

2.cd compress_yolo

3.vim Makefile and set PRUNE=1

4.start prune tiny-yolo:

./darknet detector train cfg/coco.data cfg/tiny-yolo.cfg pretrain/tiny-yolo.weights -gpus 0

5、copy backup file trained weights and test:

./darknet detector test cfg/coco.data cfg/tiny-yolo-test.cfg pretrain/tiny-yolo_prune.weights data/dog.jpg

6、test mAP:

./darknet detector valid cfg/coco.data cfg/tiny-yolo-test.cfg pretrain/tiny-yolo_prune.weights

compress the coco_results.json in results file and commit to https://competitions.codalab.org/competitions/5181#participate-submit_results

实验结果：64M compress to 18M