Kahsolt / EW-MTCNN-unofficial

Unofficial implementation of EW-MTCNN for image emotion classification

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

EW-MTCNN-unofficial

Unofficial implementation of EW-MTCNN for image emotion classification

复现论文: 基于情感轮和多任务卷积神经网络的图像情感分布学习 (2022) 赖金水,万中英,曾雪强
这个论文甚至还有后续破事水 融合情感轮注意力的情感分布学习 (2023) 陈启凡,曾雪强,左家莉,万中英,王明文;或者看 CSDN博客解读

结论:能复现出论文数据,但情感轮先验层基本上卵用没用,严重拉低精度还不如没有 🤬

⚪ 模型结构

EM-MTCNN

⚪ 情感轮先验层权重

EW_layer

训练集精度 (我懒得做五折交叉了)

⚠ inplace eval on trainset :(

-EW cls_acc ↑ sdl_kl ↓ comment
freeze 49.394% 0.0901860 冻结情感轮先验层权重
unfreeze 50.404% 0.0912241 允许情感轮先验层参与微调
none 74.495% 0.0455484 不使用情感轮先验层

references


by Armit 2023/12/11

About

Unofficial implementation of EW-MTCNN for image emotion classification

License:MIT License


Languages

Language:Python 95.7%Language:Batchfile 4.3%