Skip to main content

Notebook Tutorials

This project now includes notebook-oriented tutorials that demonstrate the framework with real datasets and end-to-end workflows.

Available notebooks​

  1. Start with the quickstart notebook to verify the environment.
  2. Use the Potsdam notebook to build a small sliding-window segmentation dataset from a real remote-sensing dataset.
  3. Finish with the inference notebook to inspect predictions and exports.

Potsdam workflow​

The Potsdam notebook is designed around a small subset of the dataset so it stays lightweight enough for local experimentation:

  1. Mount or download the Google Drive folder locally.
  2. Export a CSV with at least image and mask columns.
  3. Run build_sliding_window_dataset to create patch-level training data.
  4. Train a small segmentation model with a short schedule.
  5. Run tiled inference on a full image to validate the result.
  6. Re-run with a single tile if you want a quick smoke test.

Tutorial config​

Exact commands​

uv run python -m pytorch_segmentation_models_trainer.main \
  --config-dir pytorch_segmentation_models_trainer/conf/examples \
  --config-name potsdam_windowed_tutorial

If you want to rebuild the dataset first, use the notebook cell that calls InferenceCSVBuilder and build_sliding_window_dataset.