SDFI_ml

Machine learning framework developed and maintained by KDS for performing semantic segmentation on multichannel images.

---

Installation

Use conda or mamba (Miniforge includes conda; mamba is optional). From the repository root:

conda env create --file environment.yml
conda activate ML_sdfi
pip install --pre --no-build-isolation -r requirements_pip.txt

This installs PyTorch nightly with CUDA 12.8 (for NVIDIA Blackwell / RTX 50-series / sm_120 GPUs), fastai, git-based deps, and this package in editable mode.

Other GPUs: To use stable PyTorch instead of nightly (e.g. cu121), after the steps above run:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

(Adjust cu121 to your CUDA version; see pytorch.org/get-started/locally.)

Verify that everything works

python -c "import torch; print('CUDA available:', torch.cuda.is_available()); print('Device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'N/A')"

You should see CUDA available: True and your GPU name. If not, reinstall PyTorch with the correct CUDA index (nightly cu128 for Blackwell, or a stable cu11x/cu12x for older GPUs).

python src/ML_sdfi_fastai2/train.py --config configs/example_configs/train_example_dataset.ini

It should train for a number of epochs without errorrs

Windows

Windows: Run commands from a shell where the conda env is activated (conda activate ML_sdfi) so that Library\bin and Scripts are on PATH. After the three steps above, run once: pip install --force-reinstall pillow rasterio so PIL and rasterio use pip's Windows wheels (avoids DLL load errors when running training).

If you still see ImportError: DLL load failed while importing _imaging (or similar PIL/Pillow errors), run:

• pip install --force-reinstall pillow
• If you also see rasterio-related DLL errors: pip install --force-reinstall rasterio
  Then run your command again (e.g. train.py).

---

Usage

Training

Configuration file:
configs/example_configs/train_example_dataset.ini

Run:

python src/ML_sdfi_fastai2/train.py --config configs/example_configs/train_example_dataset.ini

---

Inference

Configuration file:
configs/example_configs/infer_example_dataset.ini

Run:

python src/ML_sdfi_fastai2/infer.py --config configs/example_configs/infer_example_dataset.ini

---

Reporting

Configuration file:
configs/example_configs/report_example_dataset.ini

Run:

In order to analyse how well the model does on a dataset we can create a .pdf report with loss plots , confusion matrixes and example iamges and predictions

cd ../multi_channel_dataset_creation
python src/multi_channel_dataset_creation/geopackage_to_label_v2.py --geopackage example_dataset/labels/example_dataset_buildings.gpkg --input_folder example_dataset/data/splitted/rgb/ --output_folder example_dataset/buildings/splitted_buildings/ --background_value 0 --value_used_for_all_polygons 1
cd ../ML_sdfi_fastai2
python src/ML_sdfi_fastai2/report.py --config configs/example_configs/report_example_dataset.ini

If you get confusing errors when creating report it is often casued by a faulty .csv file 
check ../logs_and_models/example_dataset_iter_1/iter_1/logs/iter_1.csv 
This file should look like this

epoch,train_loss,valid_loss,valid_accuracy,time,lr_0,lr_1,lr_2
0,2.671426296234131,2.261967182159424,0.15685869753360748,00:07,8e-05,8e-05,8e-05
1,2.4981231689453125,1.2677741050720215,0.6439558863639832,00:06
...
...
9,3.6438961029052734,1.2790676355361938,0.6442530751228333,00:06,8.64737438191444e-05,8.64737438191444e-05,8.64737438191444e-05

---

Example Dataset

All example configuration files are compatible with the example dataset available at:
👉 https://github.com/SDFIdk/multi_channel_dataset_creation