yolov5目标检测模型的知识蒸馏（基于响应的蒸馏）

Last update: Dec 04, 2022

Overview

代码地址：

https://github.com/Sharpiless/yolov5-knowledge-distillation

教师模型：

python train.py --weights weights/yolov5m.pt \
        --cfg models/yolov5m.yaml --data data/voc.yaml --epochs 50 \
        --batch-size 8 --device 0 --hyp data/hyp.scratch.yaml

蒸馏训练：

python train.py --weights weights/yolov5s.pt \
        --cfg models/yolov5s.yaml --data data/voc.yaml --epochs 50 \
        --batch-size 8 --device 0 --hyp data/hyp.scratch.yaml \
        --t_weights yolov5m.pt --distill

训练参数:

--weights：预训练模型

--t_weights：教师模型权重

--distill：使用知识蒸馏进行训练

--dist_loss：l2或者kl

--temperature：使用知识蒸馏时的温度

使用《Object detection at 200 Frames Per Second》中的损失

这篇文章分别对这几个损失函数做出改进，具体思路为只有当teacher network的objectness value高时，才学习bounding box坐标和class probabilities。

实验结果：

这里假设VOC2012中新增加的数据为无标签数据（2k张）。

教师模型	训练方法	蒸馏损失	P	R	mAP50
无	正常训练	不使用	0.7756	0.7115	0.7609
Yolov5l	output based	l2	0.7585	0.7198	0.7644
Yolov5l	output based	KL	0.7417	0.7207	0.7536
Yolov5m	output based	l2	0.7682	0.7436	0.7976
Yolov5m	output based	KL	0.7731	0.7313	0.7931

参数和细节正在完善，支持KL散度、L2 logits损失和Sigmoid蒸馏损失等

1. 正常训练：

2. L2蒸馏损失：

我的公众号：

关于作者

B站：https://space.bilibili.com/470550823

CSDN：https://blog.csdn.net/weixin_44936889

AI Studio：https://aistudio.baidu.com/aistudio/personalcenter/thirdview/67156

Github：https://github.com/Sharpiless

yolov5目标检测模型的知识蒸馏（基于响应的蒸馏）

Related tags

Overview

代码地址：

教师模型：

蒸馏训练：

训练参数:

实验结果：

1. 正常训练：

2. L2蒸馏损失：

我的公众号：

关于作者

Owner

yolov5目标检测模型的知识蒸馏（基于响应的蒸馏）

UFPR-ADMR-v2 Dataset

Project looking into use of autoencoder for semi-supervised learning and comparing data requirements compared to supervised learning.

Unsupervised Feature Ranking via Attribute Networks.

PipeTransformer: Automated Elastic Pipelining for Distributed Training of Large-scale Models

🚗 INGI Dakar 2K21 - Be the first one on the finish line ! 🚗

Generate images from texts. In Russian

A graph neural network (GNN) model to predict protein-protein interactions (PPI) with no sample features

Image-Scaling Attacks and Defenses

Built a deep neural network (DNN) that functions as an end-to-end machine translation pipeline

Implementation of "Debiasing Item-to-Item Recommendations With Small Annotated Datasets" (RecSys '20)

The official implementation of the Hybrid Self-Attention NEAT algorithm

Canonical Appearance Transformations

Julia and Matlab codes to simulated all problems in El-Hachem, McCue and Simpson (2021)

Source code for the NeurIPS 2021 paper "On the Second-order Convergence Properties of Random Search Methods"

PyTorch/GPU re-implementation of the paper Masked Autoencoders Are Scalable Vision Learners

Python calculations for the position of the sun and moon.

Dual Attention Network for Scene Segmentation (CVPR2019)

Public repository of the 3DV 2021 paper "Generative Zero-Shot Learning for Semantic Segmentation of 3D Point Clouds"

Pytorch implementation of our paper accepted by NeurIPS 2021 -- Revisiting Discriminator in GAN Compression: A Generator-discriminator Cooperative Compression Scheme