Using "Learning Efficient Convolutional Networks Through Network Slimming (ICCV2017) channel pruning method prune yolov5s on MSCOCO 2017 Person dataset.
MSCOCO 2017 dataset. Reserve only no crowed person class and remove all other class label.
You can download MSCOCO 2017 dataset and use scripts/create_yolov5_dataset_from_coco.py to create the dataset.
Default yolov5s.
The results are generated by test.py, Notice the speed is inconsistent with production level.
| Network | Sparsity Rate | mAPval (without finetune) 0.5 |
mAPval(finetune) 0.5 |
mAPval(finetune) 0.5:0.95 |
Speed RTX3070 (ms) |
params (M) |
|---|---|---|---|---|---|---|
| yolov5s | 62.1 | 45.2 | 3.0 | 7.0 | ||
| yolov5s 30% pruned | 1e-5 | 10.5 | 62.4 | 44.8 | 2.7 | 3.5 |
| yolov5s 50% pruned | 1e-5 | 0 | 58.8 | 41.5 | 2.3 | 1.8 |
| yolov5s 70% pruned | 1e-5 | 0 | 46.6 | 30.5 | 1.8 | 0.7 |
Set pruning ratio as 0.7 for example.
python train.py --data dataset/data.yaml --cfg models/yolov5s-person.yaml --epochs 150 --slimming --name yolov5s_coco-person_slimming --weights '' --batch-size 32 python prune_net.py --weight /path/to/the/sparsed/weight --pruning_ratio 0.7The pruned net weight and cfg file will save to "pruned_net" directory.
python train.py --data dataset/data.yaml --hyp data/hyps/hyp.finetune.yaml --cfg pruned_net/prune_0.7_pruned_net.yaml --epochs 30 --name yolov5s_coco-person_slimming_finetune_0.7 --weights pruned_net/prune_0.7_pruned_net.pt --batch-size 32 @InProceedings{Liu_2017_ICCV,
author = {Liu, Zhuang and Li, Jianguo and Shen, Zhiqiang and Huang, Gao and Yan, Shoumeng and Zhang, Changshui},
title = {Learning Efficient Convolutional Networks Through Network Slimming},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {Oct},
year = {2017}
}