Panduan Lengkap Sistem: Arsitektur, Alur & Cara Menjalankan Dokumen Tunggal Terintegrasi untuk Developer & AI Assistant | Retraining Studio & Live Counter Feedmill 1. Panduan Lengkap: Cara Menjalankan Sistem A. Menjalankan Retraining Studio dengan Docker (Direkomendasikan) cd /home/asus/feedmill/reTraining cp .env.example .env # Masukkan HF_TOKEN VIDEO_ARCHIVE_HOST=/path/ke/video/archive docker compose up -d --build # Buka Web Studio di browser: http://localhost:8080 (API di :8000) # Cek status & GPU: curl http://localhost:8000/api/health B. Menjalankan Mode Development (Manual tanpa Docker) # Terminal 1 - Backend FastAPI: cd /home/asus/feedmill/reTraining uv pip install -r requirements.txt uv pip install -e sam3/ uv run uvicorn backend.main:app --reload --port 8000 # Terminal 2 - Frontend React (Vite): cd /home/asus/feedmill/reTraining/frontend npm install npm run dev # Buka http://localhost:5173 C. Menjalankan Live Counter Lapangan (menghitung-karung) cd /home/asus/feedmill/menghitung-karung python3 predict.py # Deteksi & counter real-time python3 counter_dashboard.py # Dashboard operator di http://localhost:5000 # Atau via systemd: sudo systemctl restart karung-counter.service sudo systemctl restart karung-counter-dashboard.service D. Langkah Operasional Retraining di Web Studio 1. Projects Screen: Buat project atau pilih yang ada, upload base model (.pt), tetapkan kelas (misal: karung-pakan). 2. Video Archive: Pilih batch rekaman video per siklus kerja, tentukan rentang detik & FPS, klik Extract Frames. 3. SAM3 Auto-Label: AI otomatis mendeteksi & membuat polygon label berdasarkan teks prompt nama kelas. 4. Review Canvas: Periksa frame di web canvas, perbaiki label jika perlu, dan approve data. 5. Data Prep & Merge: Filter kualitas lalu konfirmasi merge ke Master Dataset (immutable & val split stabil). 6. YOLO Training: Fine-tuning YOLO11 dan bandingkan mAP model baru vs base model. Ambil output best.pt. 2. Alur Kerja Aplikasi (Workflow) • CCTV Stream (:8554) -> Deteksi real-time & rekaman batch MP4 di menghitung-karung. • Rekaman ditarik ke Studio Retraining -> Ekstrak frame FFmpeg -> Auto-label SAM3 -> Review Canvas. • Dataset dikunci -> YOLO Retraining -> Benchmark mAP Base vs New -> Output best.pt. 3. Penjelasan Port & Jaringan • Port 8080: Web UI Retraining Studio (Docker / Nginx) • Port 8000: FastAPI Backend (REST API & Background Jobs) • Port 5173: Frontend Dev Server (React Vite) • Port 5000: Dashboard Live Counter Lokal • Port 8554 / 554: RTSP Live Video Stream CCTV 4. Peta Filesystem & Lokasi Folder • Backend: /home/asus/feedmill/reTraining/backend/ (FastAPI, db.py, jobs) • Frontend: /home/asus/feedmill/reTraining/frontend/ (React 19, Vite 7, Tailwind) • SAM3: /home/asus/feedmill/reTraining/sam3/ (Vendor Meta SAM3) • Storage: /home/asus/feedmill/reTraining/data/ (app.db, master dataset train/val, models/best.pt) • Live Counter: /home/asus/feedmill/menghitung-karung/ (predict.py, dashboard, zones.json, jetson_counter.db) 5. Status: Before, Current, and Next • BEFORE: Model statis, pelabelan manual lambat di CVAT, tidak ada validasi mAP terstandar. • CURRENT: Retraining Studio mandiri aktif lokal di GPU workstation (SAM3 auto-label, valid split stabil, evaluasi mAP otomatis). • NEXT: Skrip otomatis ekspor TensorRT (.engine) dan fitur active learning otomatis.