Skip to content

Code

This page summarizes the organization of the released codebase and explains the role of the main scripts in the repository.

The repository contains code for:

  • dataset generation
  • dataset preprocessing and loading
  • model training
  • model evaluation
  • validation and visualization

Repository Overview

The released repository is organized around three major workflows:

  1. dataset generation

  2. construct scenes, height maps, radiomaps, and beam maps

  3. benchmark data preparation

  4. load released data into unified dense / sparse learning tasks
  5. prepare feature-map or continuous-encoding inputs

  6. define building-mask, valid-mask, and sparse-sampling logic

  7. baseline training and evaluation

  8. train and evaluate released UNet and GAN baselines

Main Script Groups

1. Dataset generation scripts

These scripts implement the released data-generation pipeline.

DatasetGeneration_Step1_OSMDownload.py

Randomly samples candidate geographic regions, filters them by building count, and downloads OSM building data.

DatasetGeneration_Step2_OSMToSionna.py

Converts downloaded OSM data into cropped 3D urban scene XML files for the downstream ray-tracing pipeline.

Environment note: this step is intended to run inside a Blender 4.0 Python environment (bpy) with Mitsuba/XML export support enabled. It is not intended to run in a standard Python interpreter.

DatasetGeneration_Step3_OSMToHeightMap.py

Rasterizes OSM building footprints into height maps for geometry-aware learning.

DatasetGeneration_Step4_RadiomapRT.py

Generates multi-beam radiomap labels by running ray tracing on scene XML files.

DatasetGeneration_Step5_RadiomapValidation.py

Provides validation and visualization utilities for checking radiomap / height-map alignment, building masks, and invalid regions.

DatasetGeneration_Step6_BeammapGenerator.py

Generates configuration-aware beam maps from a simplified LOS-based geometric model.

2. Dataset preprocessing and loading

These scripts convert the released raw dataset into benchmark-ready model inputs.

multiconfig_dataset_prepcocess_GAN.py

Dataset class and preprocessing pipeline used by the released GAN baselines.

It supports: - random / beam / scene split strategies - dense / sparse supervision modes - feature-map inputs - continuous-encoding inputs

It also defines: - building-mask extraction - valid-region masking - sparse sampling-mask generation - target normalization

multiconfig_dataset_prepcocess_Unet.py

Dataset class and preprocessing pipeline used by the released UNet baselines.

It supports: - random / beam / scene split strategies - dense / sparse supervision modes - feature-map inputs - continuous-encoding inputs - optional use of height-map geometry channels

It also defines: - building-mask extraction - valid-region masking - sparse sampling-mask generation - target normalization

3. Training scripts

These scripts train the released baselines under the benchmark settings.

ModelTraining_GAN.py

Trains GAN baselines for the released benchmark tasks.

Main supported settings: - split strategy: random / beam / scene - supervision mode: dense / sparse - condition type: feature / encoding

The training pipeline uses: - valid-region-aware adversarial loss - valid-region-aware reconstruction loss - sparse observation injection for sparse tasks

ModelTraining_Unet.py

Trains the released UNet baselines.

The implemented UNet baseline follows a two-stage WNet-style design (RadioWNet) and supports: - first_only - second_only - both

training modes.

The training pipeline supports: - feature-map inputs - continuous-encoding inputs - dense and sparse settings - optional second-stage refinement with the first-stage network frozen

4. Evaluation scripts

These scripts evaluate released pretrained checkpoints or re-trained baselines.

ModelEvaluation_GAN.py

Evaluates released GAN checkpoints and reports benchmark metrics on the test split.

ModelEvaluation_Unet.py

Evaluates released UNet checkpoints and reports benchmark metrics on the test split.

Both evaluators: - use valid-region-aware metrics - exclude building and invalid regions during the main error computation - report results under the released benchmark settings

Typical Workflow

Option A: Evaluate released checkpoints

Recommended order:

  1. prepare Dataset/
  2. prepare Pretrained_Model/
  3. run ModelEvaluation_Unet.py
  4. run ModelEvaluation_GAN.py

This is the fastest way to verify that the released resources are correctly prepared.

Option B: Retrain benchmark baselines

Recommended order:

  1. prepare Dataset/
  2. choose the target benchmark task setting
  3. use the corresponding dataset preprocessing pipeline
  4. run:
  5. ModelTraining_Unet.py, or
  6. ModelTraining_GAN.py
  7. evaluate the trained model with the corresponding evaluation script

Option C: Reproduce dataset assets

Recommended order:

  1. run DatasetGeneration_Step1_OSMDownload.py
  2. run DatasetGeneration_Step2_OSMToSionna.py
  3. run DatasetGeneration_Step3_OSMToHeightMap.py
  4. run DatasetGeneration_Step4_RadiomapRT.py
  5. optionally inspect outputs with DatasetGeneration_Step5_RadiomapValidation.py
  6. run DatasetGeneration_Step6_BeammapGenerator.py

Key Design Roles

The released codebase separates different responsibilities clearly:

Script group Main role
DatasetGeneration_* generate dataset assets
multiconfig_dataset_prepcocess_* prepare benchmark-ready model inputs
ModelTraining_* train released baselines
ModelEvaluation_* evaluate released baselines
validation / visualization tools inspect correctness and alignment

Notes

  • DatasetGeneration_Step5_RadiomapValidation.py is a utility script rather than a core data-generation step
  • DatasetGeneration_Step6_BeammapGenerator.py generates configuration-only beam maps, not ray-tracing radiomaps
  • the dataset preprocessing scripts define how building regions, invalid regions, and valid supervision masks are handled
  • sparse tasks are implemented by returning a sampling mask first, and later constructing sparse observations in the training or evaluation scripts
  • only the scene-conversion step (DatasetGeneration_Step2_OSMToSionna.py) requires a Blender 4.0 bpy environment

Related Pages

For more details, see:

  • Benchmark — task definitions and benchmark settings
  • Dataset — released data organization and label semantics
  • Pretrained Models — checkpoint organization
  • Quickstart — fastest evaluation-first workflow