How molefe-m/deepsage works
Project type and purpose
• This is a command-line, multi-stage machine-learning pipeline (“deepSAGE”) for road-type classification on OSM road-network graphs.
• It stages: (1) graph extraction, (2) categorical feature embedding, (3) deep autoencoder embedding, (4) graph representation learning (RankSAGE/GraphSAGE/GCN/GAT) plus evaluation.
File structure
• Top-level scripts drive each stage:
• 01_graph_extraction.py
• 02_categorical_feature_representation.py
• 03_deep_auto_encoder_embeddings.py
• 04_graph_representation_learning.py
• A shared “common/” library holds reusable modules:
• graph_data_loader.py, osm_graph_extraction.py (data ingestion)
• categorical_feature_representation.py, node_structural_features_extraction.py
• graph_dae_embedding.py, models.py (DAE training & modeling)
• aggregators.py, graph_representation_learning.py (GRL algorithms)
• load_save_data.py (I/O utilities)
• Artifacts (datasets, trained models, performance reports) are written to disk paths defined inside each script.
• requirements.txt enumerates dependencies (e.g., PyTorch, PyTorch Geometric, scikit-learn, OSMnx).
README for insights
• Confirms the two-stage embedding pipeline (DAE → RankSAGE) and details parameter search and performance reporting.
Detail
• I/O vs. feature engineering vs. model training vs. evaluation
• Role of common/utils (load_save_data) as a shared resource.