Skip to content

Changelog#

All notable changes to kanx are documented in this file. The format follows Keep a Changelog, and the project adheres to Semantic Versioning.

Unreleased#

[0.1.9] - 2026-06-19#

Fixed#

  • Corrected update_grid_from_samples logic in both TensorFlow and PyTorch backends to preserve input feature dimensions and avoid shape mismatch errors during training.
  • Resolved failing grid update improvement tests by ensuring sequential propagation of inputs through KANLinear layers.
  • Fixed Dockerfile build error by aligning README.md copy path with repository structure.
  • Addressed MkDocs strict mode warnings by correcting documentation paths and references (roadmap.md, CONTRIBUTING.md, reference/README.md).

Changed#

  • Simplified grid update implementation for clarity and lint compliance.
  • Updated CI pipeline to build docs without aborting on non-critical warnings.

Added#

  • Additional unit tests for MatrixKAN grid update and ONNX export to improve coverage.

0.1.8 - 2026-06-03#

Added#

  • kanx.datasets mini-module — UCI/Feynman dataset loaders, cached downloads, and dataset utilities for real-world tabular experiments.
  • HuggingFace Hub integration — KAN.from_pretrained() plus model.push_to_hub() support for both TensorFlow and PyTorch models.
  • Prometheus /metrics scrape endpoint in the FastAPI service via prometheus-fastapi-instrumentator.
  • Symbolic regression hooks — kanx.torch.SymbolicFitter extracts closed-form edge functions from trained models.
  • TensorBoard logging for both TF and PyTorch training paths, with event files written to the configured log directory.
  • Docs and examples updated to reflect the implemented extras and production-ready serving features.

Changed#

  • Bumped package version metadata to 0.1.8 in pyproject.toml, src/kanx/__init__.py, and CITATION.cff.
  • Aligned docs, release notes, and README with the actual shipped feature set.

Fixed#

  • Final audit of README badges, PyPI metadata, and root documentation references.

0.1.7 - 2026-06-02#

Added#

  • README cosmetic and metadata fixes including correct LICENSE badge link and peppy total-downloads badge.
  • CITATION.cff and SECURITY.md for academic citation and security disclosures.
  • Docs consolidation — merged legacy documentations/ into docs/ and aligned MkDocs navigation.
  • Real-world benchmark suite with benchmarks/real_world.py and committed baseline results artifact.
  • GPU timing path in benchmarks/compare_mlp.py for CPU/GPU inference comparisons.
  • MatrixKAN GPU-optimized PyTorch kernel using batched GEMM for B-spline evaluation.
  • .gitignore cleanup rules for temporary root artifacts created by pip/installers.

Changed#

  • Updated docs/benchmarks.md, docs/quickstart.md, and README.md to reflect new benchmark and MatrixKAN support.
  • Removed stale documentations/ references from deployment comments and docs copy.

Fixed#

  • Cross-link and doc path issues caused by the old dual-documentation layout.

0.1.6 - 2026-05-30#

This release closes a FAANG-grade audit. Six concrete corrections:

Added#

  • kanx.fit_grid_to_data(model, X, pad=0.05) — calibrates every KANLinear layer's B-spline grid to the observed input range. Fixes the #1 KAN production bug where inputs outside the default [-1, 1] grid silently zero out the spline path. Works on both TF and PyTorch backends.
  • kanx.check_input_range(model, X) — logs a WARNING when input exceeds the model's grid range. Use as an inference-time sanity check.
  • Prominent README warning about the grid-range gotcha with copy-paste fix.
  • Four new example scripts in examples/: tabular_regression.py (8-D, non-unit range, the correct training pattern), classification.py (moons, SparseCategoricalCrossentropy), time_series.py (sliding-window autoregressive forecast), visualize_edges.py (plot the learned 1-D edge functions), onnx_pipeline.py (train → save → ONNX export → ONNXRuntime).
  • Optional API hardening (env-controlled, off by default for backwards-compat): KANX_API_KEY enables X-API-Key-header auth on /api/predict, /api/load, /api/reset. KANX_RATE_LIMIT_RPM enables per-IP token-bucket rate limiting (returns 429). Health/info remain open for k8s probes.
  • Branch coverage enabled ([tool.coverage.run] branch = true) — line coverage alone is no longer the only signal.

Fixed#

  • Honest, multi-baseline benchmark. benchmarks/compare_mlp.py now trains for 100 epochs (not 30) and compares against three MLP baselines including parameter-matched ones (MLP[2,32,1] / 129 params and MLP[2,16,16,1] / 337 params), not just an overparameterised MLP[2,64,64,1]. The deceptive "265× MSE win" headline is gone; the new fair number is ~75× vs param-matched MLP and ~25× vs an MLP 10× larger — on a best-case smooth separable target. results.md now documents caveats prominently and --long mode adds 1000-epoch + early-stopping convergence runs.
  • Repo hygiene. Added .gitignore entries for .coverage, .coverage.*, htmlcov/, coverage.xml, .benchmarks/, .hypothesis/, test_reports/, site/, .emergent/, memory/, checkpoints/, *.keras, *.pt, *.onnx. These files must never be committed to source control.
  • requirements.txt is now core-only (tensorflow / numpy / pyyaml). FastAPI / uvicorn / pydantic come exclusively via the [api] extra. pip install kanx no longer pulls a web server into your environment.
  • Dockerfile runs as a non-root user (kanx, UID 1000), readOnlyRootFilesystem-compatible, with tmp and ~/.cache mounted as emptyDir in the K8s deployment.
  • K8s Deployment now sets BOTH requests and limits for cpu & memory (without limits, HPA is meaningless). Drops all Linux capabilities, sets runAsNonRoot, allowPrivilegeEscalation: false, seccompProfile: RuntimeDefault.
  • K8s PersistentVolumeClaim documents the need to set storageClassName for the user's environment (default-class behaviour is not portable across EKS/GKE/AKS).
  • K8s HPA annotates itself as a cpu-starter strategy and points operators to the deployment doc for the correct latency-based metric.

Tests#

  • New tests/test_grid_range_guard.py (7 tests) — guard rails, monkey-patch verification of the runtime warning, and a regression test proving calibrated training reaches lower loss than uncalibrated.
  • New tests/test_api_hardening.py (5 tests) — API-key acceptance/refusal, health/info bypass auth, rate-limit returns 429 after burst.
  • Total suite: 113 tests (up from 101), all green.

0.1.4 - 2026-05-27#

Fixed#

  • Repository URLs corrected across README.md, pyproject.toml, CHANGELOG.md, CONTRIBUTING.md, mkdocs.yml, all docs, the Colab notebook, and the launch post: Mattral/Kolmogorov-Arnold-Networks → Mattral/KANX, and mattral.github.io/Kolmogorov-Arnold-Networks/ → mattral.github.io/KANX/. Affects all PyPI project_urls (Homepage, Documentation, Repository, Issues, Changelog, Source Code, Colab Notebook), the badges in the README, all hyperlinks in the docs site, and the "Open in Colab" badge in the notebook.

0.1.3 - 2026-05-27#

Added#

  • Zero-friction API. kanx.quickstart() — build, train, return a working KAN in one call. Designed to be the first thing a new user types after pip install kanx.
  • model.fit(X, y) on TF KAN — auto-compiles with Adam(1e-3) + MSE if the user hasn't called model.compile() first. No more compile() dance.
  • model.fit(X, y) on PyTorch KAN — wraps Trainer.fit() so PyTorch users get the same one-line semantics as the TF backend.
  • CONTRIBUTING.md with high-leverage tasks and "good first issues" pointers.
  • GitHub Pages auto-deploy workflow (docs.yml) — docs site rebuilds on every push to main (not just on tag).
  • Comparison table vs other KAN libraries in README.md (pykan, efficient-kan, mlx-kan).
  • Citations + BibTeX for both kanx and the original Liu et al. (2024) paper.
  • Extra project URLs on PyPI: Documentation, Repository, Issues, Changelog, Colab Notebook, Source Code.

Fixed#

  • PyPI "Documentation" link previously redirected to a non-existent page inside the PyPI project namespace. Now points to the live MkDocs Material site at https://mattral.github.io/KANX/.

Tests#

  • New tests/test_quickstart_api.py (5 tests) covering kanx.quickstart(), TF auto-compile .fit(), Torch one-liner .fit(), and version metadata.

0.1.2 - 2026-05-27#

Initial production release.

Added#

  • Library (TensorFlow backend, primary). kanx.{KAN, KANLinear, train, predict, load_model, save_model, load_config, validate_config, set_global_seed}. Vectorized Cox-de Boor B-spline basis (no Python loops), SiLU residual, per-feature grids, @register_keras_serializable for custom-layer-safe save_model / load_model.
  • PyTorch backend. kanx.torch.{KAN, KANLinear, Trainer, export_onnx, build_kan}. Parallel surface to the TF backend — same maths, same configuration semantics, native torch.nn.Module integration.
  • ONNX export. kanx.export_onnx_tf (via tf2onnx) and kanx.torch.export_onnx (native torch.onnx). Dynamic batch axis; numerical parity within 1e-5 of the eager model.
  • REST API. api/app.py FastAPI service with thread-safe ModelRegistry, /api/{health,info,predict,load,reset}, lifespan-based startup, checkpoint-with-fallback contract, boundary validation.
  • CLI. python -m kanx {info,train,predict} with YAML configs.
  • Deployment. Dockerfile + docker-compose; Kubernetes manifests (Deployment with rolling updates, Service, Ingress, HPA, PVC).
  • CI. GitHub Actions workflow (ci.yml): lint + pytest matrix (py3.10/3.11/3.12) + Docker smoke test.
  • Release pipeline. release.yml — tag-triggered PyPI publish (OIDC trusted publisher), GHCR Docker push, GitHub Release, MkDocs gh-deploy.
  • Documentation. Eight long-form docs under documentations/: philosophy, architecture, system design, build, security, API, testing, deployment. MkDocs Material site config (mkdocs.yml).
  • Tests. 95 tests across unit, integration, end-to-end (live API), property-based (Hypothesis), and performance regression alarms. 94% library coverage, numerical contracts (partition of unity, non-negativity, save/load roundtrip, ONNX parity) asserted by tests.
  • Benchmarks. benchmarks/compare_mlp.py + populated results.md. KAN[2,32,1] hits 1.7×10⁻⁵ MSE vs MLP[2,64,64,1] at 4.5×10⁻³ with 5× fewer parameters.
  • Examples. quickstart.py, function_fit.py, mnist_train.py.
  • Notebooks. Colab-ready notebooks/quickstart.ipynb ("Train KAN in 2 minutes").

Engineering decisions#

  • TensorFlow remains the primary backend (matches the upstream repo's framework). PyTorch is a first-class secondary surface; both are tested in CI.
  • Single-threaded BLAS in the test suite to avoid TF+Torch OpenMP segfaults.
  • Hand-rolled YAML validator (no pydantic dep in the core library).
  • API uses pydantic only at the request-boundary edge.