Comparison with Other Libraries

Several Python libraries implement Self-Organizing Maps. They differ in technical architecture, maintenance, and built-in capabilities. The table below reproduces the comparison from the paper (Table 1); it is the basis for TorchSOM’s positioning.

Libraries compared: TorchSOM, MiniSom, SimpSOM, SOMPY, somoclu, and som-pbc.

Feature matrix

	TorchSOM	MiniSom	SimpSOM	SOMPY	somoclu	som-pbc
Framework	PyTorch	NumPy	NumPy	NumPy	C++	NumPy
GPU acceleration	CUDA (PyTorch)	✗	CuPy / cuML	✗	CUDA C++	✗
JIT compilation	via PyTorch	Numba [1]	✗	✗	✗ (AOT) [2]	✗
API design	scikit-learn	Custom	Custom	MATLAB	Custom	Custom
Maintenance	Active	Active	Minimal	Minimal	Minimal	✗
Documentation	Rich	Basic [3]	Basic	✗	Basic	Basic
Test coverage	90%	98%	53%	✗	Minimal	✗
PyPI distribution	✓	✓	✓	✗	✓	✗
Visualization	Advanced	✗	Moderate	Moderate	Basic	Basic
Clustering (built-in)	✓	Examples only [4]	✓	✗	✗	✗
JITL support	✓	✗	✗	✗	✗	✗
SOM variants	Multiple [5]	✗	PBC	✗	PBC	PBC
Extensibility	High	Moderate	Low	Low	Low	Low

[1]

Opt-in Numba JIT (train_batch_offline_fast). This is a compilation technique, unrelated to Just-In-Time Learning (JITL, below).

[2]

somoclu’s CUDA kernels are compiled ahead of time with nvcc when the package is built.

[3]

MiniSom ships example notebooks and partial in-code docstrings, but no narrative documentation site comparable to this one.

[4]

Clustering is not a built-in MiniSom feature; it requires user-supplied code on top of MiniSom primitives. TorchSOM provides cluster() directly.

[5]

In the current release, “Multiple” means rectangular and hexagonal topologies, each optionally toroidal via periodic boundary conditions. Growing and Hierarchical variants are on the roadmap (see the paper’s Conclusion).

Where TorchSOM fits

Existing libraries each address a specific niche: MiniSom is a minimalist, NumPy-based implementation well suited to teaching and prototyping, while somoclu targets HPC environments through CUDA C++. TorchSOM is the only library in this comparison that combines, in a single modular codebase:

• a native PyTorch backend with GPU acceleration,

• a scikit-learn-compatible API,

• an advanced built-in visualization suite,

• a built-in clustering interface,

• just-in-time-learning support, and

• multiple grid topologies with configurable neighborhood retrieval modes.

It is further supported by this narrative documentation site and a community-oriented development process, making it a complete and scalable reference for both research and production. The performance side of this comparison — comparable quantization error, lower topographic error, and faster training on most workloads — is documented in Benchmarks.

Next steps

• Benchmarks — Quantitative speed and quality comparison with MiniSom and somoclu

• Package Architecture — How TorchSOM is organized

• Quick Start — Try it yourself