Candidate Repos
작성일: 2026-05-13 · 목적: 서버 즉시 실행용 후보별 학습 config 분리
결론
권장 구조는 candidate-repos/jongno-landmark-model-candidates 단일 레포다.
레포 안에서 5개 후보를 config로 분리하고, tmux, Python venv, 6 GPU, W&B, rclone gdrive 백업을 공통 실행 계약으로 가진다.
GitHub 업로드는 owner/org와 visibility를 사용자에게 확인한 뒤 진행한다.
생성 후보
| Config | 후보 | 비교 목적 |
|---|---|---|
configs/candidates/mobileclip2_s4.yaml |
MobileCLIP2-S4 | 정확도 우선 MobileCLIP2 최상위 후보. |
configs/candidates/mobileclip2_s3.yaml |
MobileCLIP2-S3 | S4보다 가벼운 large MobileCLIP2 후보. |
configs/candidates/mobileclip2_b.yaml |
MobileCLIP2-B | 성능/비용 균형 MobileCLIP2 후보. |
configs/candidates/mobilenetv4_hybrid_large.yaml |
MobileNetV4 Hybrid Large | supervised fine-tuning image-only strong baseline. |
configs/candidates/mobilenetv4_conv_aa_large_in12k.yaml |
MobileNetV4 Conv-AA Large IN12K | ImageNet-12k pretraining 기반 MobileNetV4 accuracy-first 후보. |
데이터 전략
데이터셋은 아직 증가 중이므로, 레포에는 고정된 local split을 commit하지 않는다.
서버 실행 시점의 Dataset/<landmark_id>/labels.json과 images/를 다시 스캔해
locked test와 5-fold trainval split을 생성한다. split manifest에는 이미지 hash와 label metadata 기반
dataset_fingerprint를 남기고, W&B config에도 기록한다.
export DATA_ROOT=/data/landmark-assistant/Dataset
bash scripts/make_splits.sh
bash scripts/run_cv_tmux.sh
GitHub 업로드 전 확인
업로드에는 GitHub owner/org, public/private 여부, repo naming 확정이 필요하다.
현재 local 환경에서는 gh CLI가 PATH에 없으므로, 실제 업로드는 사용자가 owner/org를 알려준 뒤
서버 또는 GitHub CLI가 설치된 환경에서
candidate-repos/jongno-landmark-model-candidates/scripts/publish_to_github.sh를 실행한다.