Custom Model Training with GeoTrainer
Auto-labelling, loss functions built for imbalanced classes, and honest evaluation
Expert
MLOps & Pipelines
2 lessons
Samuel Appiah Kubi
About this course
Sometimes a foundation model's off-the-shelf task head isn't enough — you need a model trained specifically on your city, your crop, your building style. This course covers the full training loop: generating labels automatically instead of paying for manual annotation, choosing a loss function that actually suits imbalanced classes like buildings, and exporting the result for cheap CPU deployment.
What you'll learn
- Generate training labels automatically from OpenStreetMap and Google Open Buildings
- Explain why FocalDice loss suits building segmentation better than plain cross-entropy
- Set up GeoTrainer with mixed-precision training and checkpointing
- Export a trained model to ONNX for CPU-only production deployment
- Evaluate a segmentation model with mIoU, F1, and a confusion matrix — not just accuracy
Requirements
- Foundation Models course (recommended)
- PyTorch installed
- GPU strongly recommended (8GB+ VRAM)
Course content
Auto-Labelling from OpenStreetMap: Skipping the Manual Annotation Bill
Preview
28 min
GeoTrainer: Why FocalDice Loss, Mixed Precision, and Checkpointing All Matter
34 min
Free
- LevelExpert
- Lessons2
- CertificateYes
- AccessLifetime
Samuel Appiah Kubi
Instructor