基于YOLOv8的剪枝+知识蒸馏=无损轻量化

置顶 %海棠依旧%

已于 2025-02-09 11:46:10 修改

阅读量990

点赞数 1

文章标签： YOLO 剪枝算法 yolov8 目标检测深度学习人工智能

于 2023-10-11 18:30:55 首次发布

本文链接：https://blog.youkuaiyun.com/weixin_43175143/article/details/133766526

版权

1.实验结果（自制数据集）

（1）YOLOv8n剪枝+蒸馏（YOLOv8s为教师模型）：

Base：mAP 84.8%，Parameters 3.01M，GFLOPs 8.1G，FPS 188.7

Prune（蒸馏微调训练）：mAP 84.9%，Parameters 1.67M，GFLOPs 5.0G，FPS 217.4

（2）YOLOv8m剪枝+蒸馏（YOLOv8m自蒸馏）：

Base：mAP 96.18%，Parameters 25.84M，GFLOPs 78.692G，FPS 68.4

Prune（蒸馏微调训练）：mAP 96.34%，Parameters 6.20M，GFLOPs 37.649G，FPS 91.2

2.支持的方法

（1）剪枝方法：

l1：https://arxiv.org/abs/1608.08710

lamp：https://arxiv.org/abs/2010.07611

slim：https://arxiv.org/abs/1708.06519

group_norm：https://openaccess.thecvf.com/content/CVPR2023/html/Fang_DepGraph_Towards_Any_Structural_Pruning_CVPR_2023_paper.html

group_taylor：https://openaccess.thecvf.com/content_CVPR_2019/papers/Molchanov_Importance_Estimation_for_Neural_Network_Pruning_CVPR_2019_paper.pdf

（2）知识蒸馏：

Logits蒸馏：

（Bridging Cross-task Protocol Inconsistency for Distillation in Dense Object Detection）https://arxiv.org/pdf/2308.14286.pdf

CrossKD（Cross-Head Knowledge Distillation for Dense Object Detection）https://arxiv.org/abs/2306.11369

NKD（From Knowledge Distillation to Self-Knowledge Distillation: A Unified Approach with Normalized Loss and Customized Soft Labels）https://arxiv.org/abs/2303.13005

DKD(Decoupled Knowledge Distillation) https://arxiv.org/pdf/2203.08679.pdf

LD(Localization Distillation for Dense Object Detection) https://arxiv.org/abs/2102.12252

WSLD(Rethinking the Soft Label of Knowledge Extraction: A Bias-Balance Perspective) https://arxiv.org/pdf/2102.00650.pdf

Distilling the Knowledge in a Neural Network https://arxiv.org/pdf/1503.02531.pd3f

RKD(Relational Knowledge Disitllation) http://arxiv.org/pdf/1904.05068。

特征蒸馏：

CWD（Channel-wise Knowledge Distillation for Dense Prediction）https://arxiv.org/pdf/2011.13256.pdf

MGD（Masked Generative Distillation）https://arxiv.org/abs/2205.01529

FGD（Focal and Global Knowledge Distillation for Detectors）https://arxiv.org/abs/2111.11837

FSP（A Gift from Knowledge Distillation: Fast Optimization, Network Minimization and Transfer Learning）https://openaccess.thecvf.com/content_cvpr_2017/papers/Yim_A_Gift_From_CVPR_2017_paper.pdf

PKD（General Distillation Framework for Object Detectors via Pearson Correlation Coefﬁcient） https://arxiv.org/abs/2207.02039

VID(Variational Information Distillation for Knowledge Transfer) https://arxiv.org/pdf/1904.05835.pdf

Mimic（Quantization Mimic Towards Very Tiny CNN for Object Detection）https://arxiv.org/abs/1805.02152

代码复现不易，有偿获取。