CIFAR-10 で 95% → 97% に押し上げる「効くやつ」5点。Label smoothing + 強い augmentation (Cutout / Mixup / CutMix) + 推論時 TTA。
5 problems. Solve each template in place, then run the last cell check("...") to grade. Stuck? Open the linked solution in the rightmost column.
| Order | # | Problem | Difficulty | Solution | Colab |
|---|---|---|---|---|---|
| 1 | 51 | 51_label_smoothing_ce.ipynb — Label Smoothing Cross-Entropy |
Easy | ↗ |  |
| 2 | 45 | 45_cutout.ipynb — Cutout / RandomErasing |
Medium | ↗ | |
| 3 | 46 | 46_mixup.ipynb — Mixup |
Medium | ↗ | |
| 4 | 47 | 47_cutmix.ipynb — CutMix |
Medium | ↗ | |
| 5 | 48 | 48_tta_hflip.ipynb — TTA (Horizontal Flip Averaging) |
Easy | ↗ | |
# from the repo root
make run # or `docker compose up`
# then in JupyterLab, navigate to practice/W5/ and open a notebookIf you accidentally break a template, delete it and rerun python scripts/build_weeks.py to restore — it only copies missing files, so your other in-progress notebooks are preserved.