windows-pyqt-ultralytics-desktop-app

name	windows-pyqt-ultralytics-desktop-app
description	Use when a Windows project already has YOLO or Ultralytics weights, configs, local custom modules, or inference code and the user wants a runnable PyQt5 desktop detection app, PyCharm-ready package, batch/image/video/camera workflow, or clean launcher without retraining.

Windows PyQt Ultralytics Desktop App

Core Workflow

1. Inspect the real project assets before writing UI code:
   • weights such as best.pt or last.pt,
   • model or dataset YAML files,
   • local custom modules,
   • existing detect.py, predict.py, or UI entrypoints.
2. Reuse existing inference logic before building a new detector stack.
3. Register local custom Ultralytics modules before loading checkpoints.
4. Make Windows startup stable:
   • import Torch/Ultralytics before PyQt when DLL load order is sensitive,
   • use path-safe OpenCV read/write on Chinese or non-ASCII paths,
   • set the working directory from Path(__file__).resolve().parent.
5. Build the smallest complete GUI that matches the requested workflow:
   • image selection,
   • video selection,
   • camera mode when requested,
   • folder batch mode,
   • result preview,
   • progress/log panel,
   • saved output folder.
6. Add threaded workers for video or batch work, with cooperative pause/resume when requested.
7. Provide a PyCharm/debug path and a clean Windows launch path.
8. Smoke test with the actual shipped assets or included sample files.

Decision Rules

• Do not retrain when usable weights already exist and the user asked for an app.
• Do not replace local model code until checkpoint loading requirements are understood.
• Avoid broad rewrites when an existing detection function can be wrapped.
• Package sample images or a smoke-test path so the delivered app does not depend on hidden local data.
• Use pythonw.exe, .pyw, or a shortcut only for the no-console launch path; keep python main.py for debugging.

Pitfalls

• Checkpoint load fails because custom modules are not registered under the names expected by the checkpoint.
• PyQt starts before Torch and causes DLL or WinError 1114 failures.
• OpenCV fails only on Chinese-path images.
• A PowerShell smoke test uses Bash-style heredoc syntax; use PowerShell here-strings or a script file instead.
• Pause/resume processes one extra item because the worker does not yield after emitting progress.

Verification

• The app loads the actual local checkpoint/config.
• Required custom modules are registered before model load.
• GUI startup works on Windows.
• Image read/write works for the user's path style.
• Requested controls exist and process at least one sample input.
• Batch or video output saves to the expected folder.
• Pause/resume, if implemented, stops and resumes from the intended point.