티스토리 뷰

cool
https://github.com/huggingface/candle/pull/3405
fix: conv2d_tiled produces wrong results when C == H == W by developer0hye · Pull Request #3405 · huggingface/candle
Bug conv2d_tiled produces completely wrong output when C_in == H == W (e.g., input shape [1, 64, 64, 64]). Fixes #3404 Root Cause conv2d_tiled converts input from NCHW to NHWC internally. It uses a...
github.com
이거 다 클로드로 한거다.
문제 발견이랑 검증 과정 설계만 직접하고 구현은 다 클로드한테 맡김
pr 에 description 그 누가 읽어도 납득가게 쓰라고 하고, merged pr 참고해서 description 톤, 방식 지키라고하고 또 repo내에 contribution guide, code formatting guide 있으면 따르라고 말함. 또 이런 건 엄밀하게 정의되는 test case가 있어야하니 test case 설계잘하라고 시킴
https://github.com/developer0hye/candle-birefnet
GitHub - developer0hye/candle-birefnet: BiRefNet inference for Hugging Face Candle
BiRefNet inference for Hugging Face Candle. Contribute to developer0hye/candle-birefnet development by creating an account on GitHub.
github.com
이거 구현하다가 발견한 문제였다.
아무리 요새 linear, attention layer 가 크게 중요하다고 하지만 vision 분야의 근본 연산인 conv연산에 오류가 있다니!! 말이 안 되지 않는가. 이렇게 오픈소스는 에러가 많고 기여할 여지가 굉장히 많다.
'Deep Learning' 카테고리의 다른 글
| WSL 에서 Claude 설치시 안 된다면 (0) | 2026.03.03 |
|---|---|
| 구글에서 작성한 프롬프트 샘플 및 설계 전략 (0) | 2026.01.11 |
| Qwen3VL 예제 코드 모음 (1) | 2026.01.06 |
| Object Detection 에 적합한 백본 - RF DETR 논문에서 발췌 (0) | 2025.12.28 |
| 20251122 요새 써보고 있는 논문 요약 프롬프트 (0) | 2025.11.22 |
- Total
- Today
- Yesterday
- 백트래킹
- 백준 11053
- Lowest Common Ancestor
- 자료구조
- ㅂ
- 순열
- MOT
- FairMOT
- C++ Deploy
- 가장 긴 증가하는 부분 수열
- LCA
- 파이참
- 조합
- 문제집
- 백준 11437
- 백준 1766
- 단축키
- cosine
- 백준
- 위상 정렬 알고리즘
- 인공지능을 위한 선형대수
- 이분탐색
- PyCharm
| 일 | 월 | 화 | 수 | 목 | 금 | 토 |
|---|---|---|---|---|---|---|
| 1 | 2 | 3 | 4 | 5 | 6 | 7 |
| 8 | 9 | 10 | 11 | 12 | 13 | 14 |
| 15 | 16 | 17 | 18 | 19 | 20 | 21 |
| 22 | 23 | 24 | 25 | 26 | 27 | 28 |
| 29 | 30 | 31 |
