# Self-supervised image retrieval

`ssl_retrieval` trains retrieval representations for the 64×64 unlabeled image dataset. It contains frozen DINO inference, SimCLR, an autoencoder pretext task, four-way rotation prediction, and joint training. Retrieval always uses the normalized encoder embedding and cosine similarity.

## Commands

```bash
python -m ssl_retrieval train --method simclr --output checkpoints/simclr.pt
python -m ssl_retrieval train --method joint --output checkpoints/joint.pt
python -m ssl_retrieval evaluate --method joint --checkpoint checkpoints/joint.pt
python -m ssl_retrieval retrieve --method joint --checkpoint checkpoints/joint.pt \
  --gallery data/dataset2/public_test/public_test/gallery.json \
  --queries data/dataset2/public_test/public_test/queries.json \
  --output submission_public.csv
```

Use `--from-scratch` to disable the default ImageNet initialization. Joint loss weights are configurable with `--contrastive-weight`, `--reconstruction-weight`, and `--rotation-weight`.

Training reserves a deterministic pseudo-gallery of 1,000 train images and creates 200 degraded queries from it. The remaining 2,000 images are used for optimization. The best checkpoint maximizes the mean of Recall@1, Recall@5, Recall@10, and mAP@10. Defaults target one 8 GB GPU and the competition's 20-minute limit; reduce `--epochs` if needed.