[AI Tech 3기 Level 2 P Stage] Data Annotation

ConVinsight 🧑‍💻

Convenience + Insight : 이용자의 편의를 찾는 통찰력

Member

Wrap Up Report 📑

💻 ConVinsight level2-data-annotation Notion

Final Score 🏆

• Public f1 score 0.7089 → Private f1 score 0.6790
• Public 6위 → Private 7위

Competition Process 🗓️

Time Line

Project Outline

batch size 실험

• [x] batch size 12
• [x] batch size 16
• [x] batch size 24
• [x] batch size 32

Data Augmentation 실험

• [x] Blur
• [x] Color
• [x] Inverting
• [x] Outline
• [x] Rotate

Optimizer 실험

• [x] SGD (lr = 1e-3)
• [x] Adam (lr = 1e-3)
• [x] AdamW (lr = 1e-3)
• [x] SGD (lr = 1e-2)

Learning Rate Scheduler 실험

• [x] MultiStepLR
• [x] CosineAnnealingWarmupRestarts
• [x] CosineAnnealingLR

Folder Structure 📂

level2-data-annotation_cv-level2-cv-05/
│
├── 📂 baseline
│   ├── 📝 convert_mlt.py
│   ├── 📝 scheduler.py
│   ├── 📝 train.py
│   └── 📝 inference.py
│
└── 📂 utils
    ├── 📂 Convert_Custom
    │    ├── 📝 from_doc_to_ufo.ipynb
    │    └── 📝 from_doc_to_ufo.py
    ├── 📂 UFO_visualization
    │    └── 📝 data_visualization.ipynb
    └── 📂 multi_label_stratified_kfold