Skip to content

Latest commit

 

History

History
67 lines (49 loc) · 2.57 KB

RESULTS.md

File metadata and controls

67 lines (49 loc) · 2.57 KB
Raw

Challenge

  • 现存模型过拟合
  • VA这种连续空间难以理解

Model

hparam: E=100, B=32, lr=2e-5

baseline

Accuracy (train/valid) ResNet50 ResNet101 MobileNet_V2 ViT_B_16 ViT_B_32
TwitterI 98.71%/83.00% 98.74%/82.76% 97.6%/80.16% 98.58%/82.55% 99.16%/79.06%
EmoSet 89.79%/76.74% 82.18%/77.42% 76.10%/74.93% 92.06%/78.20% 83.55%/74.64%
Artphoto 97.83%/38.89% 98.68%/35.77% 96.13%/29.65% 99.34%/38.26% 97.52%/28.55%
Abstract 98.74%/18.40% 99.57%/14.57% 92.46%/18.71% 98.78%/15.00% 99.73%/19.04%
Emo6Dim7 97.00%/39.00% 98.56%/42.24% 89.75%/45.56% 99.61%43.80% 98.04%/40.42%
Emo6Dim6 97.95%/45.72% 97.21%/45.54% 92.05%/47.17% 98.26%/49.93% 98.26%/44.80%
Emo6VA 0.188/1.135 0.17/1.12 0.2081/0.7786 0.03781/0.6638 0.04957/0.7252
OASIS 0.07928/0.4739 0.5141/0.4202 0.179/0.5316 0.02315/0.4182 0.07642/0.6762

别人论文或者项目代码里抄的

Accuracy xxx xxx xxx xxx xxx
xxx

ours

⚪ Head=linear (数值精度可能低一些，但可以看出相关性)

run vis.py => img/tx/*.png 可视化权重矩阵/相关性 run vis_gui.py => 交互式可视化

Head Dataset M-ResNet50 M-MobileNet_V2
Polar TwitterI 90.26%/83.62% 86.62%/77.24%
Mikels EmoSet 75.16%/67.01% 72.76%/70.80%
EkmanN Emo6Dim7 50.79%/41.91% 46.50%/41.68%
Ekman Emo6Dim6 54.72%/48.36% 50.31%/46.64%
VA Emo6VA 0.5276/0.6186 0.5401/0.6707

⚪ Head=mlp (数值精度可能高一些，但无法透视相关性)

run vis_gui.py => 交互式可视化

Head Dataset M-ResNet50 M-MobileNet_V2 M-ViT_B_16
Polar TwitterI 92.45%/84.70% 86.10%/80.94% 91.68%/81.93%
Mikels EmoSet 75.39%/72.46% 76.58%/70.63% 76.39%/68.86%
EkmanN Emo6Dim7 52.33%/43.35% 44.49%/42.02% 52.72%/41.15%
Ekman Emo6Dim6 57.13%/45.96% 48.88%/47.44% 55.59%/43.14%
VA Emo6VA 0.4911/0.6186 0.5664/0.6691 0.4923/0.6303

Analysis

  • 多任务模型 (9 datasets, 3/5 model archs, 2 head types)
    • 精度是否更好，最好略好
    • 有没有减缓过拟合
    • 在精度不下降太多的情况下，一定程度上节省了参数
    • LDL六种度量
  • 空间转换
    • 可视化权重矩阵/相关性: run vis.py => img/tx/*.png
    • 交互式可视化: run vis_gui.py