chipoint-2 / RUN_GUIDE.md
4latte
chipoint-2: street-view + general-photo arms, organized
31ac821
|
Raw
History Blame Contribute Delete
2.36 kB
# RUN GUIDE β€” chipoint-2 street-view + general-photo arms
## Files
- `street_gallery/proj_head_gallery_512_{dinov3,so400m256,gopt}.npy` β€” street gallery
projection vectors (4.89M OSV-5M street frames, unit-norm 512-d fp16,
the street-arm index)
- `street_gallery/gallery_all_C_so400m256.npy` β€” street gallery coordinates (4.89M x 2)
- `weights/proj_head_{dinov3,so400m256,gopt}.pt` β€” projection heads (tower -> 512-d)
- `weights/streetview_k512_joint.pt` β€” street-view reranker arm (K512, 23 features)
- `photo_gallery/` β€” unified photo gallery, 14,541,898 rows (MP14 10.9M +
MP16-Pro 4.1M): `emb_*.npy` (512-d fp16 projected), `ll_*.npy` (lat/lon),
`ids_*.txt` (source ids, row-aligned), `manifest.json`
- `weights/photoarm_k768_thr.pt` / `weights/photoarm_k768_sr.pt` / `weights/photoarm_k512_thr.pt` β€”
general-photo reranker arms (23 features)
- `weights/ghead_merged_z4/z6/z8/z10.pt`, `weights/ghead_merged_reg.pt`,
`features/gallery_merged_cells.npz` β€” merged-head feature stack (required for the
23-feature contract)
- `weights/router_osv_flickr.pt` β€” street/photo router (DINOv3 512-d -> logit;
>0 photo, else street)
## Load
```bash
python run_chipoint.py weights/streetview_k512_joint.pt # street arm: verify + schema
python run_chipoint.py weights/photoarm_k768_thr.pt # photo arm: verify + schema
```
Checkpoint dict: `state` (SR), `thr_state` (ThrNet), `mu`/`sd`
(normalizers), `nf`, `names`, `thr_km`.
## Feature contract (nf = 23)
18 base features over the candidate pool (per-encoder sims, margins,
ranks, cross-encoder agreement, vote counts, spatial density) + 5
merged-head features from `ghead_merged_z4/z6/z8/z10/reg.pt` mapped by
`gallery_merged_cells.npz`. Feed order is `ckpt['names']`; standardize with
`(X - mu) / sd` before scoring.
## Eval protocol (numbers in README)
Street arm: OSV-5M 4.89M gallery; 210,122-image sequence-stratified test
split. Retrieval: query tower embedding -> `proj_head_*` -> 512-d, top-K
against `street_gallery/proj_head_gallery_512_*` (same tower space), rerank with the
SR/ThrNet head on the 23-feature stack, threshold at 1/25/200/750/2500 km.
Photo arm: unified 14.5M gallery (`photo_gallery/emb_*.npy`, 512-d fp16);
im2gps3k (2997) / yfcc4k (4536) queries; blocklisted honest protocol
(184 exact-id leaks removed); 5-fold CV for the headline numbers.