Instructions for HW2P2 (Youngin Kim, youngin2)

11-785 HW2P2: Face Recognition & Face Verification

0. Prerquisite

Please modify configs/path/custom.yaml for your own path

You can see Hyperparameters values in configs/config.yaml.

Or You can see the details and experimental results in WanDB.

- Recognition best model: convnext-4772-16:13:41:23
- Verification best model: resnet50-4662-softtriple-sgd0.01-ls0.1-15:19:40:04

Model Architecture
- You can check in models/convnext.py and resnet.py
- I change the block number [3,3,9,3] to [4,7,7,2] / [4,6,6,2] because local feature is important in this task.
- However, in the verification task, custom resnet model is better than convnext.
Optimizer
- AdamW
- scheduler: CosineAnnealing
Regularize
- weight decay
Augmentation
- You can check in datasets/transform.py
TTA
- TTA results in slightly higher scores (+0.001)
- You can check in tta.ipynb

$ python run.py save_name={name_for_submission&weight_file}

Name		Name	Last commit message	Last commit date
Latest commit History 12 Commits
bash		bash
configs		configs
datasets		datasets
models		models
utils		utils
.gitignore		.gitignore
HW2P2_Starter.ipynb		HW2P2_Starter.ipynb
README.md		README.md
debug.ipynb		debug.ipynb
kaggle_submission.ipynb		kaggle_submission.ipynb
run.py		run.py
tmp.ipynb		tmp.ipynb
triplet.ipynb		triplet.ipynb
tta copy.ipynb		tta copy.ipynb
tta.ipynb		tta.ipynb