Run registry · 2026-06-12
실험 실행 기록과 결과 반영 템플릿
모든 run을 같은 형식으로 남기기 위한 문서다. 좋은 결과만 남기지 않고, screening, main, failed, excluded run을 분리해 기록한다.
기록 원칙
결과가 좋았는지보다 먼저, 어떤 질문에 답하기 위한 run인지와 비교 가능한 조건인지 확인한다. 비교 조건이 맞지 않으면 결과가 좋아도 main conclusion으로 쓰지 않는다.
Run Category
| Category | 의미 | 논문에서 쓰는 방식 |
|---|---|---|
main |
사전에 정의한 fair comparison matrix에 속하는 run. | Results와 Ablation Study의 핵심 표로 사용한다. |
screening |
후보 탐색, 빠른 비교, 조건이 완전히 맞지 않는 초기 실험. | 실험 설계 변경의 근거로만 사용한다. |
failed |
학습 실패, export 실패, accuracy collapse, 환경 문제. | 실패 실험/한계/다음 개선 근거로 기록한다. |
excluded |
run은 끝났지만 비교 조건이 맞지 않아 최종 표에서 제외한 결과. | 왜 제외했는지 Method 또는 Appendix에서 설명한다. |
Run Record Template
Run:
category: main | screening | failed | excluded
run_name:
wandb_project:
wandb_run_id:
local_run_dir:
status: queued | running | completed | failed | excluded
Question:
research_question:
hypothesis:
changed_variable:
controlled_variables:
Dataset:
root:
fingerprint:
split_manifest:
fold:
class_count:
confirmed_records:
Config:
path:
backbone:
train_method:
image_unfreeze_ratio:
text_unfreeze_ratio:
classification_loss:
contrastive_loss:
hard_negative_loss:
epochs:
image_size:
batch_size:
seed:
Result:
val_top1:
test_top1:
test_top3:
macro_f1:
hard_case_top1:
low_margin_count:
text_retrieval_r1:
Artifacts:
checkpoint:
onnx:
int8:
checksum:
Interpretation:
conclusion:
comparison_scope:
failure_or_exclusion_reason:
next_step:
현재 Registry Snapshot
아래 표는 현재 문서 기준의 registry snapshot이다. live W&B 상태가 아니라 문서화된 상태이므로, 새 export를 받으면 갱신한다.
| Run group | Category | 상태 | 해석 | 문서 |
|---|---|---|---|---|
| S3 pilot 5-fold | screening |
completed | 이름은 server_full이었지만 실제 config는 partial unfreeze였으므로 최종 S3/S4 결론에는 사용하지 않는다. | screening result |
| S4 partial ArcFace 5-fold | screening |
completed | S3 pilot과 조건이 달라 직접 우열 결론은 보류한다. | screening result |
| S3/S4 2x4 paper matrix | main |
running / queued | 공정 비교의 source of truth. 완료 후 최종 backbone/train method 결론을 작성한다. | paper matrix |
| Sprint 1 AI Hub w8a16 INT8 | failed |
recorded | NPU latency는 좋았지만 embedding collapse가 있어 배포 정확도 후보가 아니다. | mobile artifact benchmark |
결과 그래프 작성 규칙
- main 결과는 fold별 점과 mean/std를 함께 보인다.
- Top-1만 막대로 보이지 말고 Top-3, macro-F1, hard-case를 함께 둔다.
- 차이가 작은 경우 y축 확대 그래프와 0-1 전체 범위 그래프를 구분한다.
- screening 결과와 main 결과는 같은 색상/범례로 섞지 않는다.
- deployment 그래프는 정확도 그래프와 분리한다. 예: file size, latency, memory.
Error Analysis Template
| 오류 유형 | 예시 | 확인할 로그 | 다음 조치 |
|---|---|---|---|
| 궁궐 문루/전각 혼동 | 광화문, 근정문, 근정전 | confusion matrix, low-margin csv, top-3 prediction | parent/sub landmark 정책, confusion_prior, hard negative 보강. |
| 배경 landmark 혼동 | 광화문 이미지 후경의 청와대 | top-3 score, background_landmark_ids, image crop | 주 피사체 기준 라벨링, ambiguity policy, crop robustness 평가. |
| 저품질 입력 | 흐림, 어두움, tiny upscaled, noise | quality detector score, confidence status | 앱단 image quality gate와 low_quality policy 개선. |
| 자연어 검색 실패 | 한국어 묘사형 query | text retrieval ranking, alias/keyword score, fusion score | catalog query_examples, bilingual caption, text encoder fine-tuning 검토. |
결과 반영 체크리스트
- run category가 명확한가?
- 비교 조건이 공정한가?
- dataset fingerprint와 split manifest가 적혔는가?
- fold별 결과와 mean/std가 모두 있는가?
- 성공 결과와 실패/제외 사유가 모두 남았는가?
- 그래프가 어떤 값을 보여주는지 축과 범례가 명확한가?
- 결론이 결과보다 과장되어 있지 않은가?
- 다음 실험 또는 앱 artifact handoff가 명시되어 있는가?