How michelin/torchsom works
Project type and purpose
• Open-source Python library (“tool”) for building and visualizing Self-Organizing Maps (SOMs).
• Leverages PyTorch for GPU acceleration and seamless integration into ML pipelines.
File structure
• Top-level modules:
• torchsom/configs – model and visualization configuration
• torchsom/core – core SOM implementations (batch, growing, hierarchical)
• torchsom/utils – helper functions (distances, decay, initialization, metrics, grid, topology)
• torchsom/visualization – plotting and map-drawing utilities
• docs, notebooks, data, tests, CI/CD scripts
• External points: PyTorch, Jupyter notebooks, PyPI packaging.
README for insights
• Emphasis on GPU support, scalability, modular design.
Main components and their interactions
a. User Code (scripts or notebooks)
b. TorchSOM API
• Config Loader (torchsom/configs)
• SOM Engine (torchsom/core)
• Utilities (torchsom/utils)
• Visualizer (torchsom/visualization)
c. External Dependencies
• PyTorch (CPU/GPU computation)
• Jupyter (interactive demos)
• CI/CD (GitHub Actions)
• PyPI (distribution)
d. Data
• Input datasets (e.g., iris.csv)
• Output artifacts (plots, logs, model weights)
Architectural patterns and design principles
• Modular layering (Configs → Core → Utils → Visualization)
• Clear API boundary between user code and low-level SOM operations
• Separation of concerns: computation vs. configuration vs. plotting