尚未偵測到更新。GitHub 發布版本後會顯示在這裡。
依最近已完成的 8 小時觀測回合間的 GitHub 星號增量排序
個人 GitHub 帳號下公開儲存庫的擁有者可以登入並驗證所有權，自行管理介紹，無需管理員核准。
登入並產生驗證檔案，在 30 分鐘內將其提交至此儲存庫的預設分支。其他分支或分叉儲存庫中的檔案無效。然後在下方檢查檔案，即可無需管理員核准直接管理專案。請保留檔案以維持管理權限。
如需登入，將前往登入頁面。登入後請回到此處，再次點擊所選的驗證按鈕。
在專案資料夾中執行 /nomorevibe verify，即可驗證所有權並自行管理介紹。
一分鐘完成驗證🌸 High-accuracy botanical vision pipeline — real-time species classification leveraging cloud-hosted ViT models, low-latency edge proxying, and client-side web acceleration.
翻譯準備中
取自儲存庫 README
正在載入完整 README。 以下為已儲存的摘錄。
Point your camera at a flower. Know what it is in seconds.
Flower identification for 116 species, powered by a fine-tuned Vision Transformer.
Ever seen a flower on a walk and had no idea what it was called?
Snap a photo, upload one, or paste an image link, and get a species name plus a
confidence score in a few seconds. No app to install, no account needed to try it.
📸 Snap or upload a photo of any flower
🔒 Sent over HTTPS via our inference proxy
🌼 ViT identifies it 116 species, top-5 scores
📚 Learn more pulled live from Wikipedia檢視完整 README ↗Flower identification for 116 species, powered by a fine-tuned Vision Transformer.
Ever seen a flower on a walk and had no idea what it was called?
Snap a photo, upload one, or paste an image link, and get a species name plus a confidence score in a few seconds. No app to install, no account needed to try it.
Identification runs server-side on a fine-tuned ViT-base. Your image is sent over HTTPS to do that, used only to run the prediction, and not stored — see the privacy page for specifics.
[!NOTE] Earlier betas ran a smaller CNN in-browser via TensorFlow.js, so images stayed on your device. The ViT is far more accurate but too large to ship to a browser, so inference moved to a server. The privacy tradeoff is deliberate and documented rather than glossed over.
browser ──image──▶ Cloudflare Worker ──image + X-Proxy-Secret──▶ ONNX backend
(origin allowlist, (FastAPI +
holds the secret) onnxruntime)
◀── { flower, confidence, top_k } ────────────────────────────────────┘
The Worker keeps the backend from being an open public API: it enforces an origin allowlist and attaches a shared secret the backend checks. Only labels and scores come back, so the weights never leave the container.
| Path | What lives there |
|---|---|
*.html, *.js, favicon.svg | The static frontend (GitHub Pages) |
space/ | FastAPI + onnxruntime inference server |
proxy/ | Cloudflare Worker that fronts the backend |
training/ | Kaggle training + ONNX conversion scripts |
tools/ | Build script for the Trefle species data |
Only the frontend files are published to the web root —
.github/workflows/pages.yml copies those into _site/ and deploys that, so the
training scripts stay in version control without being served.
Frontend only — it talks to the deployed Worker, so there's no build step:
git clone https://github.com/gsor56/findflower.git
cd findflower
python -m http.server 8000 # http://localhost:8000 is on the CORS allowlist
The backend and Worker each have their own setup notes in
space/README.md and proxy/README.md.
Got a misidentification? Hit "No, incorrect" right on the try page — every bit of feedback helps retrain the model.
This is a solo, student-built project — feedback, bug reports, and suggestions are genuinely welcome. Open an issue or reach out via the contact page.
Released under the MIT License — free to use, modify, and learn from.