anofit
양품 사진만으로 학습하는 기계 시각 이상 탐지. 긁힘·깨짐 같은 구조적 불량과, 부품이 빠졌거나 자리가 바뀐 논리적 불량을 함께 잡는다. 학습부터 임계값 결정, 검사 시험까지 대시보드 하나로 한다.
Anomaly detection for machine vision, trained from good images only — structural and logical defects, with a dashboard. Documentation is in Korean; the CLI prints Korean.
- 문서·이슈: https://github.com/pashidl-lab/anofit — 대시보드 따라하기:
docs/dashboard-walkthrough.md - 문의·라이선스: pashidl.lab@gmail.com
무엇을 하나
| 모드 | 잡는 것 | 바탕 |
|---|---|---|
struct |
긁힘 · 깨짐 · 오염 · 변형 | SALAD (ICCV 2025) — 교사/학생 + 오토인코더 |
logic |
부품 누락 · 개수 오류 · 위치 바뀜 | CSAD (BMVC 2024) — 부품 분할 + 히스토그램 |
dpat |
양품과 다른 국소 패턴 (0.1.1) | DINOv2 패치 특징 + 양품 메모리 (학습 없음) |
dhist |
있어야 할 것이 없음 — 작은 부품 누락 (0.1.2) | DINOv2 패치를 64개 시각 단어로 묶어 단어별 개수 (학습 없음) |
both |
둘 다 (기본) | 점수를 z-score 로 합침 |
0.1.1 부터는 어느 분기를 쓸지 품종마다 정한다 — 티칭 때 받은 불량이 8장 이상이면 그것으로
고르고(후보 넷 {struct, +dhist, +dpat, +dpat+dhist} 중, 기본값보다 2%p 이상 나을 때만 교체), 없으면
구조부 기본값을 쓴다. 5품종 실측: 고정 both 90.91 · 고정 구조부 92.03 · 불량으로 고르면 93.00
(breakfast_box 는 85.71 → 94.04). 논리부(phist)는 그대로 학습되며 임계값 탭에서 both 로 고를 수 있다.
MVTec LOCO 5품종 실측 (AUROC, 양품 = test/good, 이상 = logical + structural, 2026-09-09):
| 품종 | struct | logic | both |
|---|---|---|---|
| juice_bottle | 99.90 | 73.27 | 99.72 |
| breakfast_box | 86.52 | 70.30 | 84.97 |
| pushpins | 92.66 | 53.23 | 91.57 |
| screw_bag | 82.17 | 49.96 | 72.94 |
| splicing_connectors | 95.26 | 62.07 | 94.34 |
이 숫자는 라벨 없이, 양품만 보고 학습한 값이다. 논문 값은 라벨된 테스트 셋으로 가장 좋은
에폭을 고른 것이라 직접 비교가 안 된다. 자세한 것은 저장소의 BENCH_REPORT.md.
설치
# 1) CUDA 에 맞는 torch 를 먼저 (pip 은 CPU 판을 고른다)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121
# 2) anofit — 학습까지 하려면 [train]
pip install "anofit[train]"
# 검사만 하는 기계(이미 학습한 번들로 점수만)면 이것으로 충분하다
pip install anofit
- Python 3.10 – 3.13, Windows / Linux
- 학습은 GPU 가 필요하다 (실측 요구 VRAM 은
anofit doctor가 말해 준다). 검사는 CPU 로도 된다. - 산출물·캐시·가중치를 어느 디스크에 둘지는
anofit.yaml로 정한다 — 본보기anofit.example.yaml. 없으면 사용자 캐시 폴더에 쌓인다.
5분 시작
anofit doctor --mode both # 이 기계로 되는가: GPU · 디스크 · 의존성 · 가중치
anofit fetch --mode both # 파운데이션 가중치 10.4 GB (원 배포처에서, sha256 대조, 이어받기)
anofit serve # 대시보드 → http://127.0.0.1:8732
대시보드에서: 준비 탭에 양품 폴더(서버 경로 또는 업로드) → 학습 탭에서 품종 이름을 정하고 시작 → 끝나면 임계값 탭에서 운용점을 고르고 검사 탭에 사진을 넣어 본다.
명령줄로 같은 일:
anofit train --category my_part --good ./good --mode both # 번들 생성
anofit threshold <번들> # 임계값이 뜻하는 것
anofit score <번들> photo.png --models-dir <번들> # 한 장 판정
anofit export <번들> -o my_part.afz # 현장으로 옮길 파일 하나
가중치는 동봉하지 않는다
anofit fetch 가 원 배포처에서 받는다. 각 가중치는 배포처의 라이선스를 따른다
(THIRD_PARTY_NOTICES.md §6, 매니페스트 anofit/fetch_manifest.json).
| 크기 | 출처 | |
|---|---|---|
| GroundingDINO Swin-T | 0.69 GB | IDEA-Research |
| RAM++ Swin-L | 3.01 GB | recognize-anything |
| SAM-HQ ViT-H | 2.57 GB | SysCV |
| SAM ViT-H | 2.56 GB | Meta |
| EfficientAD teacher | 0.03 GB | SALAD 저장소 |
| Imagenette v2 | 1.56 GB | fast.ai |
--mode struct 만이면 6.7 GB, logic 만이면 6.3 GB. 첫 학습 때 라이브러리가 스스로 받는 것이
1 GB 쯤 더 있다 (bert-base-uncased, timm wide_resnet50_2, dino_vitbase8).
라이선스 · 체험
평가용 라이선스다 (LICENSE). 처음 실행한 날부터 90일은 그대로 쓴다. 그 뒤에는
학습·대시보드·fetch 가 멈추고, 이미 학습한 번들로 검사하는 것은 계속 된다 — 돌고 있는
라인이 서지 않는다.
anofit license # 남은 날짜
anofit license --install "<key>" # 받은 키 설치 (또는 환경 변수 ANOFIT_LICENSE)
키는 pashidl.lab@gmail.com 으로 요청한다. 학습한 모델과 사진은 전부 사용자의 것이다.
포함된 오픈소스(GroundingDINO · SAM · RAM++ · CSAD · SALAD)는 각자의 라이선스를 그대로
가지며, 무엇을 고쳤는지는 THIRD_PARTY_NOTICES.md 에 파일 단위로 적었다.
안 되는 것 · 알아 둘 것
- 안전이 걸린 곳에는 쓰지 않는다 (
LICENSE§7). - 정확도는 사진·조명·부품에 달렸다. 위 표는 MVTec LOCO 에서의 측정이지 약속이 아니다.
- 학습 시간: 품종당 최대 2시간 (RTX 4090, both, 양품 300장 기준). 캐시 1~3 GB, 번들 0.1 GB.
- Windows 에서 폴더 가져오기는 심볼릭 링크를 쓰고, 안 되면 복사한다.
Release files for anofit 0.1.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Built distributions (wheels)
Total release size: 5.7 MB
Release files / anofit-0.1.2-cp313-cp313-win_amd64.whl
| Download URL | anofit-0.1.2-cp313-cp313-win_amd64.whl |
|---|---|
| Size | 694.7 kB |
| Tags | CPython 3.13 Windows x86-64 |
|
SHA-256 checksum How to use checksums |
0063e91d2cde58e7f98e53e445a01fc2fff734ede47be14df11289f72d591e30
|
|
BLAKE2b-256 checksum How to use checksums |
fe2f81ebe54c3ff5f62691f99f029d4c68ca7f10ea1e716754c8504dd16dd8cf
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl
| Download URL | anofit-0.1.2-cp313-cp313-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 728.4 kB |
| Tags | CPython 3.13 Linux glibc 2.17+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
4ac209e92b6017edd8d7044feddf7d48c19c8fe3185af2a1fbf04255e82a1893
|
|
BLAKE2b-256 checksum How to use checksums |
7ee014807a1f2dfba4dc69b51ccb1ff09856af220a30315266fee7adfd7cc364
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp312-cp312-win_amd64.whl
| Download URL | anofit-0.1.2-cp312-cp312-win_amd64.whl |
|---|---|
| Size | 695.9 kB |
| Tags | CPython 3.12 Windows x86-64 |
|
SHA-256 checksum How to use checksums |
983c2203070068e7f7ed6bd68e2a374d859e34c2ae1efd6537ff81eb31ed30b9
|
|
BLAKE2b-256 checksum How to use checksums |
cf32d0450703b55b6f2847917db0e878a041b97f91b0bbd293e3a083a343e3e8
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl
| Download URL | anofit-0.1.2-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 727.1 kB |
| Tags | CPython 3.12 Linux glibc 2.17+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
dacb45ac5cf666f0a1a3aaa4891a0990376f6fddfec4ea7079cfe2865f2acf53
|
|
BLAKE2b-256 checksum How to use checksums |
a980cb23e23d4422d707f8f1a2afa93b78728920dbb81ef94b7b013fb48a341e
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp311-cp311-win_amd64.whl
| Download URL | anofit-0.1.2-cp311-cp311-win_amd64.whl |
|---|---|
| Size | 705.6 kB |
| Tags | CPython 3.11 Windows x86-64 |
|
SHA-256 checksum How to use checksums |
1527e19e022883d703c905d636f289516ba84c28dcfe6f4863e5b229b81e87e0
|
|
BLAKE2b-256 checksum How to use checksums |
889a30a4a781d6112d8fb41bb0b9af9b518bd3ae900b973ce56deae52fb385eb
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl
| Download URL | anofit-0.1.2-cp311-cp311-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 729.1 kB |
| Tags | CPython 3.11 Linux glibc 2.17+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
4c2178644097e2b22e54c6a36ea2b508f5390517aec2a75d0ab94c3a556c5d4a
|
|
BLAKE2b-256 checksum How to use checksums |
2dd510be00c3bfe0e4bfb7a877e8745b532426318fffe135c7b35ad0319f5eed
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp310-cp310-win_amd64.whl
| Download URL | anofit-0.1.2-cp310-cp310-win_amd64.whl |
|---|---|
| Size | 705.2 kB |
| Tags | CPython 3.10 Windows x86-64 |
|
SHA-256 checksum How to use checksums |
a02116396fd4110efb9560c26c12b6aaaf6c05ab6b924ffaee53a7e95d86177b
|
|
BLAKE2b-256 checksum How to use checksums |
0f6603a89aa1682ba373462fef13019956b6f952f3f22f51c0e2785bc386d073
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency logRelease files / anofit-0.1.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl
| Download URL | anofit-0.1.2-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.manylinux_2_28_x86_64.whl |
|---|---|
| Size | 728.8 kB |
| Tags | CPython 3.10 Linux glibc 2.17+ x86-64 Linux glibc 2.28+ x86-64 |
|
SHA-256 checksum How to use checksums |
be91a667ecf9e039de3c82c555cc76b32d58449f342f26c396d664d470b1b1f7
|
|
BLAKE2b-256 checksum How to use checksums |
26258c07df78408506966a0fcfc14c1b3387f52be657acf350253aa8527c4a2d
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 18, 2026.
Transparency log