How romaincorbel/airfrans works
Project Type and Purpose
• Type: A research-oriented Python ML framework for training and evaluating graph/neural models on the AirfRANS CFD dataset.
• Purpose: Ingest CFD data, train multiple model architectures (MLP, GraphSAGE, PointNet, GUNet) on various tasks, compute metrics, store scores, and produce visualizations.
File Structure
• Top-level scripts
• main.py: CLI entry for training, scoring
• train.py: orchestration of training loops
• dataset.py, naca_generator.py, reorganize.py: data ingestion & preprocessing
• metrics.py, metrics_NACA.py: metric computations & logging
• get_results.py: collate scores from JSON into summaries
• params.yaml: hyperparameters & task configurations
• Directories
• models/: implementations of the four architectures
• metrics/: training logs and plots generated during training
• scores/: JSON score outputs for each task (full, scarce, reynolds, aoa)
• notebooks (.ipynb): dataset_stats.ipynb for data exploration, visualization.ipynb for post-hoc plots
README Insights
• CLI usage patterns (main.py arguments)
• Required external dataset folder “Dataset/”
• Output locations: metrics/ for training logs, scores/TASK/ for evaluation results, notebooks for stats and visualizations
Main Components & Their Relationships
a. CLI & Config
• User invokes main.py with arguments + params.yaml
b. Data Ingestion & Preprocessing
• naca_generator.py: synthesize NACA airfoil parameter sets
• reorganize.py: restructure raw dataset
• dataset.py: PyTorch Dataset object that loads data from Dataset/
c. Model Layer
• models/MLP.py, GraphSAGE.py, PointNet.py, GUNet.py: define forward pass & architecture
d. Training Engine
• train.py: batches data, calls model, backpropagates, logs metrics via metrics.py
e. Metrics & Logging
• metrics.py & metrics_NACA.py: compute losses, force-coefficient metrics
• metrics/ folder: store logs, plots per epoch
f. Evaluation & Scoring
• main.py with “-s 1” or get_results.py: compute test-set metrics, write JSON in scores/TASK/
g. Visualization & Analysis
• visualization.ipynb: generates plots of force coefficients, boundary layers
• dataset_stats.ipynb: generates basic dataset statistics
Architectural Patterns & Principles
• Modular separation: data / model / training / evaluation / visualization
• Config-driven: params.yaml lets you tweak hyperparameters and task type
• Extensible: adding a new model means dropping its .py into models/ and invoking via CLI
• File-based pipelines: input raw data → preprocess → training → metric logs → score JSON → notebook visualization