Model weights - L2 global + L5 Kerala

Pretrained checkpoints for KPlaceNet (L2 global model and L5 Kerala pinpoint mode).

These are not stored in git (GitHub blocks files over 100 MB). Download into checkpoints/ with:

python scripts/download_weights.py

Manual mapping (asset -> destination):

• l2_imagenet_1.0_layer4_last.pt -> checkpoints/l2_imagenet_1.0_layer4/last.pt (global demo model; also the L5 backbone warm-start source)
• l5_kerala_kmeans300_last.pt -> checkpoints/l5_kerala_kmeans300/last.pt (Kerala model used by the demo)
• l5_kerala_kmeans300_best.pt -> checkpoints/l5_kerala_kmeans300/best.pt (best validation checkpoint)
• l2_cells_full.json -> checkpoints/l2_cells_full.json (full 300 global cell boxes for the map)
• l3_uncertainty_results.json -> checkpoints/l3_uncertainty_results.json (temperature + conformal quantile used by the demo)
• l5_kerala_kmeans300_cells.json -> checkpoints/l5_kerala_kmeans300/cells.json (300 Kerala cells)
• l5_kerala_kmeans300_metrics.json -> checkpoints/l5_kerala_kmeans300/metrics_l5_kerala.json
• l5_kerala_kmeans300_results.json -> checkpoints/l5_kerala_kmeans300/results.json

Datasets are not included. Re-create the Kerala slice with scripts/extract_kerala_osv5m.py and the global test subset with scripts/download_subset.py.