https://github.com/bigtuo/yolov3-channel-and-layer-pruning

yolov3 yolov4 channel and layer pruning, Knowledge Distillation 层剪枝，通道剪枝，知识蒸馏

https://github.com/bigtuo/yolov3-channel-and-layer-pruning

Repository

yolov3 yolov4 channel and layer pruning, Knowledge Distillation 层剪枝，通道剪枝，知识蒸馏

https://github.com/Bigtuo/yolov3-channel-and-layer-pruning/blob/master/

# yolov3-channel-and-layer-pruning
[ultralytics/yolov3](https://github.com/ultralytics/yolov3)[Learning Efficient Convolutional Networks Through Network Slimming (ICCV 2017)](http://openaccess.thecvf.com/content_iccv_2017/html/Liu_Learning_Efficient_Convolutional_ICCV_2017_paper.html)bnGmma



yolov3bn gamma[](https://github.com/zbyuan)



![](https://github.com/tanluren/yolov3-channel-and-layer-pruning/blob/master/data/img/1.jpg)




####  
1.**yolov3-spp**yolov3-spp.weightsyolov3

2.train.py,  --img_size 608 .

3.2019/12/06

4.2019/12/08********HintonDistilling the Knowledge in a Neural Networkcfg--t_cfg  --t_weightsyolo

5.2019/12/10734912150 

6.2019/12/14[apex](https://github.com/NVIDIA/apex),train.pymixed_precisionFalse.

7.2019/12/23****"Learning Efficient Object Detection Models with Knowledge Distillation"targetL2targetcfg

8.2019/12/27****

9.2020/01/02testimg_sizetest

10.2020/01/04[](https://blog.csdn.net/weixin_41397123/article/details/103828931)**visdrone**12M

11.2020/04/10**yolov3-tiny**slim_prune.py

12.2020/4/24**yolov4**.

13.2020/4/30datasets.py 592.

14.2020/7/8**yolov4-tiny**.




#### 
requirements.txt[](https://github.com/ultralytics/yolov3/wiki/Train-Custom-Data)darknet

yolov3cfgdatayolov3.weights



`python train.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/yolov3.weights --epochs 100 --batch-size 32`

#### 
scale0.001mAP,BNs;-sr--prune 0prune.py--prune 1 ss

ptepoch`python -c "from models import *; convert('cfg/yolov3.cfg', 'weights/last.pt')"`darknet weightsepochdarknet weightsepoch 0



`python train.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/last.weights --epochs 300 --batch-size 32 -sr --s 0.001 --prune 1`
* ##### s
sss



`bn_module.weight.grad.data.add_(s * torch.sign(bn_module.weight.data))`
* ##### s
epochs0.5s1000.5sbns0.5100bnsssprune_utils.pyBNOptimizer



`# s = s if epoch <= opt.epochs * 0.5 else s * 0.01`
* ##### s
epochs0.585%15%s10085%baselinestrain.py0.850.5100



`#if opt.sr and opt.prune==1 and epoch > opt.epochs * 0.5:`

`#  idx2mask = get_mask2(model, prune_idx, 0.85)`

#### 
[Lam1360/YOLOv3-model-pruning](https://github.com/Lam1360/YOLOv3-model-pruning)yolov323shortcutaddshortcutLam1360/YOLOv3-model-pruningshortcut



`python prune.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/last.pt --percent 0.85`

#### 
[coldlarry/YOLOv3-complete-pruning](https://github.com/coldlarry/YOLOv3-complete-pruning)shortcutshortcutmaskmaskshortcutshortcut



`python shortcut_prune.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/last.pt --percent 0.6`

#### 
[PengyiZhang/SlimYOLOv3](https://github.com/PengyiZhang/SlimYOLOv3)maskshortcutmaskmergemask



`python slim_prune.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/last.pt --global_percent 0.8 --layer_keep 0.01`

#### 
shortcutCBLGmmayolov3shortcutshortcutshortcutyolov323shortcut8shortcut2416shortcut4869shortcut



`python layer_prune.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/last.pt --shortcuts 12`

#### 




`python layer_channel_prune.py --cfg cfg/my_cfg.cfg --data data/my_data.data --weights weights/last.pt --shortcuts 12 --global_percent 0.8 --layer_keep 0.1`

#### finetune
6epochwarmup



`python train.py --cfg cfg/prune_0.85_my_cfg.cfg --data data/my_data.data --weights weights/prune_0.85_last.weights --epochs 100 --batch-size 32`

#### tensorboard
`tensorboard --logdir runs`



![tensorboard](https://github.com/tanluren/yolov3-channel-and-layer-pruning/blob/master/data/img/2.jpg)


issue734912150



#### 
yolov3-sppoxfordhand[](http://www.robots.ox.ac.uk/~vgg/data/hands/downloads/hand_dataset.tar.gz),dataconverter.pytrain.txtvalid.txtoxfordhand.data

`python train.py --cfg cfg/yolov3-spp-hand.cfg --data data/oxfordhand.data --weights weights/yolov3-spp.weights --batch-size 20 --epochs 100`



`python -c "from models import *; convert('cfg/yolov3.cfg', 'weights/last.pt')"`

`python train.py --cfg cfg/yolov3-spp-hand.cfg --data data/oxfordhand.data --weights weights/converted.weights --batch-size 20 --epochs 300 -sr --s 0.001 --prune 1`



100epochbaselinemAP0.84;s0.001300epochprune 1shortcutepochs0.70.9Gmma0.1mAP 0.797

![baseline_and_sparse](https://github.com/tanluren/yolov3-channel-and-layer-pruning/blob/master/data/img/baseline_and_sparse.jpg)


bntensorboardbnGmmaGmma1Gmma00

![bn](https://github.com/tanluren/yolov3-channel-and-layer-pruning/blob/master/data/img/bn.jpg)


layer_channel_prune.pyslim_prunelayer_pruneGlobal perent0.93layer keep0.01shortcuts1648(32CBL16shortcut)239M5.2MmAP0.534.6M98%166tesla p100

`python layer_channel_prune.py --cfg cfg/yolov3-spp-hand.cfg --data data/oxfordhand.data --weights weights/last.pt --global_percent 0.93 --layer_keep 0.01 --shortcuts 16`



![prune9316](https://github.com/tanluren/yolov3-channel-and-layer-pruning/blob/master/data/img/prune9316.png)


50epoch0.793bnbaseline10shortcut50epoch0.810.95542.8M5mAP050epoch0.75



`python train.py --cfg cfg/prune_16_shortcut_prune_0.93_keep_0.01_yolov3-spp-hand.cfg --data data/oxfordhand.data --weights weights/prune_16_shortcut_prune_0.93_keep_0.01_last.weights --batch-size 52 --epochs 50`

![finetune_and_bn](https://github.com/tanluren/yolov3-channel-and-layer-pruning/blob/master/data/img/finetune_and_bn.jpg)

cfgyolov3cfgweights

[](https://pan.baidu.com/s/1APUfwO4L69u28Wt9gFNAYw)cfgbaseline

## License
Apache 2.0

Owner

GitHub Events