A desktop GUI application for training and running inference with Detectron2 models.
- Train object detection and instance segmentation models with a visual interface
- Live monitoring — real-time loss plot and training logs
- Inference on single images or entire folders with result visualization
- Model management — save and load trained weights for later use
| Model | Task |
|---|---|
| Faster R-CNN (R50-FPN, R101-FPN) | Object Detection |
| RetinaNet (R50-FPN, R101-FPN) | Object Detection |
| Mask R-CNN (R50-FPN, R101-FPN) | Instance Segmentation |
# Install dependencies (see INSTALL.md for detailed setup)
pip install -r requirements.txt
# Run the application
python main.pyAlchemyDetect uses COCO JSON format for training datasets. You need:
- A directory containing your training images
- A COCO-format JSON annotation file
- Open the Train tab
- Select your training images directory and COCO JSON annotation file
- Choose a model architecture from the dropdown
- Set hyperparameters (learning rate, iterations, batch size)
- Choose an output directory
- Click Start Training
- Monitor progress via the log viewer and loss plot
- Open the Inference tab
- Click Load Model and select a trained
.pthfile (config.yaml will be auto-detected if in the same directory) - Adjust the confidence threshold
- Click Run on Image or Run on Folder
- Browse results using the navigation buttons
- Python 3.10 or 3.11
- PyQt6 — Desktop GUI
- Detectron2 — Object detection / instance segmentation
- PyTorch — Deep learning backend
- pyqtgraph — Real-time loss plotting