prosemble 2.1.1__tar.gz → 2.1.2__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {prosemble-2.1.1 → prosemble-2.1.2}/.github/workflows/ci.yml +1 -1
- {prosemble-2.1.1 → prosemble-2.1.2}/PKG-INFO +1 -1
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/distance.py +7 -12
- prosemble-2.1.2/prosemble/core/kernel.py +222 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/losses.py +5 -5
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/similarities.py +2 -2
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/__init__.py +15 -0
- prosemble-2.1.2/prosemble/models/dk_glvq.py +246 -0
- prosemble-2.1.2/prosemble/models/dk_heskes_som.py +230 -0
- prosemble-2.1.2/prosemble/models/dk_kohonen_som.py +219 -0
- prosemble-2.1.2/prosemble/models/dk_matrix_lvq.py +248 -0
- prosemble-2.1.2/prosemble/models/dk_neural_gas.py +174 -0
- prosemble-2.1.2/prosemble/models/dk_relevance_lvq.py +274 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble.egg-info/PKG-INFO +1 -1
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble.egg-info/SOURCES.txt +8 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/pyproject.toml +1 -1
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/api/models.rst +33 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/conf.py +1 -1
- prosemble-2.1.2/sphinx-docs/guides/differentiating_kernels.rst +294 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/index.rst +1 -0
- prosemble-2.1.2/tests/test_dk_models.py +379 -0
- prosemble-2.1.1/prosemble/core/kernel.py +0 -103
- {prosemble-2.1.1 → prosemble-2.1.2}/.bumpversion.cfg +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.editorconfig +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.envrc +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.github/ISSUE_TEMPLATE/config.yml +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.github/ISSUE_TEMPLATE/feature_request.md +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.github/workflows/release.yml +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.gitignore +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/.readthedocs.yaml +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/CONSTITUTION.md +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/LICENSE +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/MANIFEST.in +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/ONNX_COVERAGE.md +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/README.md +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/TASK.md +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/benchmarks/benchmark_distance_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/benchmarks/benchmark_fcm_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/benchmarks/benchmark_pcm_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/devenv.lock +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/devenv.nix +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/devenv.yaml +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/afcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/bgpc_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/cbc_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/celvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/fcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/fpcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/glvq_advanced_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/glvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/glvq_unequal_prototypes_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/gmlvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/grlvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/hcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/heskes_som_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/ipcm2_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/ipcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kafcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kfcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kfpcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kipcm2_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kipcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kmeans_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/knn_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kohonen_som_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kpcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/kpfcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/large_dataset_training.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/lcelvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/lmrslvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/lmrslvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/lvq1_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/lvqmln_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/mcelvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/mrslvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/mrslvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/neural_gas_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/npc_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/oc_lmrslvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/oc_lmrslvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/oc_mrslvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/oc_mrslvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/oc_rslvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/oc_rslvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/pcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/pfcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/plvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/rslvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/slng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/smng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/som_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/srng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/stng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/tcelvq_ng_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/vis_fcm_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/vis_glvq_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/vis_neural_gas_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/examples/vis_som_iris.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/flake.nix +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/mkdocs.yml +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/nix/package.nix +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/__init__.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/__init__.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/activations.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/callbacks.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/competitions.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/data.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/distributed.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/initializers.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/manifolds.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/model_selection.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/onnx_export.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/pipeline.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/pooling.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/protocols.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/quantization.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/serialization.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/utils.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/vis.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/core/visualizer.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/datasets/__init__.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/datasets/dataset.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/afcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/base.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/bgpc.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/cbc.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/celvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/celvq_ng_mixin.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/crossentropy_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/fcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/fpcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/glvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/growing_neural_gas.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/hcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/heskes_som.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/image_cbc.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/image_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/ipcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/ipcm2.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kafcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kfcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kfpcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kipcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kipcm2.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kmeans.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/knn.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kohonen_som.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kpcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/kpfcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/lcelvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/local_matrix_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/lvq1.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/lvq21.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/lvqmln.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/matrix_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/matrix_rslvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/mcelvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/median_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/mrslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/neural_gas.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/npc.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_glvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_glvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_glvq_ng_mixin.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_gmlvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_gmlvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_grlvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_grlvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_gtlvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_gtlvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_lgmlvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_lgmlvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_mrslvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_mrslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_rslvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_rslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/oc_rslvq_ng_mixin.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/pcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/pfcm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/plvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/probabilistic_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/prototype_base.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/relevance_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/riemannian_neural_gas.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/riemannian_slng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/riemannian_smng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/riemannian_srng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/riemannian_stng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/rslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/siamese_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/slng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/smng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/som.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/srng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/stng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/svq_occ.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/svq_occ_lm.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/svq_occ_m.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/svq_occ_r.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/svq_occ_t.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/tangent_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/models/tcelvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble/py.typed +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble.egg-info/dependency_links.txt +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble.egg-info/requires.txt +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/prosemble.egg-info/top_level.txt +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/requirements_docs.txt +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/setup.cfg +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/Makefile +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/api/core.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/advanced.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/clustering.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/one_class.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/onnx.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/riemannian.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/supervised.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/guides/topology.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/installation.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/make.bat +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/sphinx-docs/quickstart.rst +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/__init__.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_activations.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_cbc.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_celvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_checkpoint.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_clustering_patience.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_competitions.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_core.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_core_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_data.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_dataset.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_dataset_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_distance_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_distributed.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_examples.sh +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_fcm_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_glvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_gmlvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_heskes_som.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_image_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_initializers.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_jax_models.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_jit_inference.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_kohonen_som.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_lcelvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_losses.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_lvq_nongradient.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_lvqmln_plvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_matrix_rslvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_mcelvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_minibatch.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_mixed_precision.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_model_selection.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_mrslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_neural_gas.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_oc_glvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_oc_glvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_oc_glvq_variants.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_oc_mrslvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_oc_rslvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_oc_rslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_onnx_export.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_pcm_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_pfcm_jax.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_pipeline.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_pooling.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_probabilistic.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_protocols.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_prototype_base.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_quantization.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_riemannian_neural_gas.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_riemannian_supervised_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_rslvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_scan_training.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_serialization.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_siamese_lvq.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_similarities.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_slng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_smng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_srng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_stng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_svq_occ.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_svq_occ_r.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_svq_occ_variants.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_tcelvq_ng.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/tests/test_vis.py +0 -0
- {prosemble-2.1.1 → prosemble-2.1.2}/uv.lock +0 -0
|
@@ -33,7 +33,7 @@ jobs:
|
|
|
33
33
|
python -m pip install --upgrade pip
|
|
34
34
|
pip install .[all]
|
|
35
35
|
- name: Run tests
|
|
36
|
-
run: pytest tests/ -x -q --tb=short --ignore=tests/test_oc_glvq.py --ignore=tests/test_oc_glvq_ng.py --ignore=tests/test_oc_glvq_variants.py --ignore=tests/test_svq_occ.py --ignore=tests/test_svq_occ_r.py --ignore=tests/test_svq_occ_variants.py --ignore=tests/test_smng.py --ignore=tests/test_slng.py --ignore=tests/test_stng.py --ignore=tests/test_riemannian_neural_gas.py
|
|
36
|
+
run: pytest tests/ -x -q --tb=short --ignore=tests/test_oc_glvq.py --ignore=tests/test_oc_glvq_ng.py --ignore=tests/test_oc_glvq_variants.py --ignore=tests/test_svq_occ.py --ignore=tests/test_svq_occ_r.py --ignore=tests/test_svq_occ_variants.py --ignore=tests/test_smng.py --ignore=tests/test_slng.py --ignore=tests/test_stng.py --ignore=tests/test_riemannian_neural_gas.py --ignore=tests/test_riemannian_supervised_ng.py
|
|
37
37
|
|
|
38
38
|
build:
|
|
39
39
|
name: Build package
|
|
@@ -380,17 +380,12 @@ def tangent_distance_matrix(
|
|
|
380
380
|
"""
|
|
381
381
|
Compute pairwise localized tangent distances.
|
|
382
382
|
|
|
383
|
-
Each prototype j has an orthogonal subspace basis Omega_j of shape (d, s).
|
|
383
|
+
Each prototype j has an orthogonal subspace basis :math:`\Omega_j` of shape (d, s).
|
|
384
384
|
The tangent distance projects out the subspace directions:
|
|
385
385
|
|
|
386
|
-
|
|
386
|
+
.. math::
|
|
387
387
|
|
|
388
|
-
|
|
389
|
-
diff = x - w_j
|
|
390
|
-
proj = Omega_j^T @ diff (project onto subspace)
|
|
391
|
-
recon = Omega_j @ proj (reconstruct in ambient space)
|
|
392
|
-
tangent_diff = diff - recon (residual orthogonal to subspace)
|
|
393
|
-
d = ||tangent_diff||^2
|
|
388
|
+
d(x, w_j) = \|(I - \Omega_j \Omega_j^T)(x - w_j)\|^2
|
|
394
389
|
|
|
395
390
|
Parameters
|
|
396
391
|
----------
|
|
@@ -484,9 +479,9 @@ def gaussian_kernel_matrix(
|
|
|
484
479
|
True
|
|
485
480
|
|
|
486
481
|
Kernel Trick:
|
|
487
|
-
For feature map phi mapping to infinite-dimensional Hilbert space,
|
|
488
|
-
K(x, y) =
|
|
489
|
-
|
|
482
|
+
For feature map :math:`\phi` mapping to infinite-dimensional Hilbert space,
|
|
483
|
+
:math:`K(x, y) = \langle\phi(x), \phi(y)\rangle`. Kernel distance:
|
|
484
|
+
:math:`\|\phi(x) - \phi(y)\|^2 = K(x,x) - 2K(x,y) + K(y,y) = 2 - 2K(x,y)` for normalized kernel.
|
|
490
485
|
|
|
491
486
|
Use Cases:
|
|
492
487
|
- Kernel Fuzzy C-Means (KFCM)
|
|
@@ -643,7 +638,7 @@ def omega_distance(x: chex.Array, y: chex.Array, omega: chex.Array) -> chex.Arra
|
|
|
643
638
|
"""
|
|
644
639
|
Omega (projection-based) distance between two vectors.
|
|
645
640
|
|
|
646
|
-
Computes
|
|
641
|
+
Computes :math:`\|\text{diff} \cdot \Omega\|^2` where :math:`\text{diff} = x - y`.
|
|
647
642
|
|
|
648
643
|
Args:
|
|
649
644
|
x: Array of shape (d,)
|
|
@@ -0,0 +1,222 @@
|
|
|
1
|
+
"""
|
|
2
|
+
JAX-based kernel functions for kernel clustering methods.
|
|
3
|
+
|
|
4
|
+
This module provides GPU-accelerated kernel computations using JAX.
|
|
5
|
+
"""
|
|
6
|
+
|
|
7
|
+
import jax.numpy as jnp
|
|
8
|
+
import chex
|
|
9
|
+
from jax import jit
|
|
10
|
+
from functools import partial
|
|
11
|
+
|
|
12
|
+
|
|
13
|
+
@jit
|
|
14
|
+
def gaussian_kernel(x: chex.Array, y: chex.Array, sigma: float) -> chex.Array:
|
|
15
|
+
"""
|
|
16
|
+
Compute Gaussian (RBF) kernel between two vectors.
|
|
17
|
+
|
|
18
|
+
:math:`K(x, y) = \exp(-\|x - y\|^2 / (2\sigma^2))`
|
|
19
|
+
|
|
20
|
+
Args:
|
|
21
|
+
x: First vector, shape (n_features,)
|
|
22
|
+
y: Second vector, shape (n_features,)
|
|
23
|
+
sigma: Kernel bandwidth parameter
|
|
24
|
+
|
|
25
|
+
Returns:
|
|
26
|
+
Kernel value (scalar)
|
|
27
|
+
"""
|
|
28
|
+
diff = x - y
|
|
29
|
+
sq_norm = jnp.sum(diff * diff)
|
|
30
|
+
return jnp.exp(-sq_norm / (2.0 * sigma ** 2))
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
@jit
|
|
34
|
+
def batch_gaussian_kernel(
|
|
35
|
+
X: chex.Array, Y: chex.Array, sigma: float
|
|
36
|
+
) -> chex.Array:
|
|
37
|
+
"""
|
|
38
|
+
Compute Gaussian kernel between two sets of vectors.
|
|
39
|
+
|
|
40
|
+
Args:
|
|
41
|
+
X: First set of vectors, shape (n_samples, n_features)
|
|
42
|
+
Y: Second set of vectors, shape (m_samples, n_features)
|
|
43
|
+
sigma: Kernel bandwidth parameter
|
|
44
|
+
|
|
45
|
+
Returns:
|
|
46
|
+
Kernel matrix, shape (n_samples, m_samples)
|
|
47
|
+
K[i, j] = K(X[i], Y[j])
|
|
48
|
+
"""
|
|
49
|
+
# Compute squared Euclidean distances
|
|
50
|
+
X_sq = jnp.sum(X ** 2, axis=1, keepdims=True) # (n_samples, 1)
|
|
51
|
+
Y_sq = jnp.sum(Y ** 2, axis=1, keepdims=True) # (m_samples, 1)
|
|
52
|
+
|
|
53
|
+
# ||x - y||^2 = ||x||^2 + ||y||^2 - 2<x, y>
|
|
54
|
+
sq_distances = X_sq + Y_sq.T - 2.0 * (X @ Y.T) # (n_samples, m_samples)
|
|
55
|
+
|
|
56
|
+
# Ensure non-negative (numerical stability)
|
|
57
|
+
sq_distances = jnp.maximum(sq_distances, 0.0)
|
|
58
|
+
|
|
59
|
+
# Compute kernel
|
|
60
|
+
K = jnp.exp(-sq_distances / (2.0 * sigma ** 2))
|
|
61
|
+
|
|
62
|
+
return K
|
|
63
|
+
|
|
64
|
+
|
|
65
|
+
@jit
|
|
66
|
+
def kernel_distance_squared(
|
|
67
|
+
X: chex.Array, Y: chex.Array, sigma: float
|
|
68
|
+
) -> chex.Array:
|
|
69
|
+
"""
|
|
70
|
+
Compute squared distance in feature space.
|
|
71
|
+
|
|
72
|
+
For Gaussian kernel: :math:`\|\phi(x) - \phi(y)\|^2 = K(x,x) + K(y,y) - 2K(x,y) = 2(1 - K(x,y))`,
|
|
73
|
+
since :math:`K(x,x) = 1`.
|
|
74
|
+
|
|
75
|
+
Args:
|
|
76
|
+
X: First set of vectors, shape (n_samples, n_features)
|
|
77
|
+
Y: Second set of vectors, shape (m_samples, n_features)
|
|
78
|
+
sigma: Kernel bandwidth parameter
|
|
79
|
+
|
|
80
|
+
Returns:
|
|
81
|
+
Squared distances in feature space, shape (n_samples, m_samples)
|
|
82
|
+
"""
|
|
83
|
+
K = batch_gaussian_kernel(X, Y, sigma)
|
|
84
|
+
return 2.0 * (1.0 - K)
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
@jit
|
|
88
|
+
def kernel_distance(
|
|
89
|
+
X: chex.Array, Y: chex.Array, sigma: float
|
|
90
|
+
) -> chex.Array:
|
|
91
|
+
"""
|
|
92
|
+
Compute distance in feature space.
|
|
93
|
+
|
|
94
|
+
Args:
|
|
95
|
+
X: First set of vectors, shape (n_samples, n_features)
|
|
96
|
+
Y: Second set of vectors, shape (m_samples, n_features)
|
|
97
|
+
sigma: Kernel bandwidth parameter
|
|
98
|
+
|
|
99
|
+
Returns:
|
|
100
|
+
Distances in feature space, shape (n_samples, m_samples)
|
|
101
|
+
"""
|
|
102
|
+
D_sq = kernel_distance_squared(X, Y, sigma)
|
|
103
|
+
return jnp.sqrt(jnp.maximum(D_sq, 0.0))
|
|
104
|
+
|
|
105
|
+
|
|
106
|
+
@jit
|
|
107
|
+
def kernel_distance_squared_per_proto(
|
|
108
|
+
X: chex.Array, W: chex.Array, sigmas: chex.Array
|
|
109
|
+
) -> chex.Array:
|
|
110
|
+
"""
|
|
111
|
+
Squared kernel distance with per-prototype bandwidth.
|
|
112
|
+
|
|
113
|
+
.. math::
|
|
114
|
+
|
|
115
|
+
d_\\kappa^2(x, w_k) = 2\\left(1 - \\exp\\left(-\\frac{\\|x - w_k\\|^2}{2\\sigma_k^2}\\right)\\right)
|
|
116
|
+
|
|
117
|
+
Each prototype :math:`w_k` has its own bandwidth :math:`\\sigma_k`.
|
|
118
|
+
|
|
119
|
+
Args:
|
|
120
|
+
X: Data matrix, shape (n_samples, n_features).
|
|
121
|
+
W: Prototype matrix, shape (n_prototypes, n_features).
|
|
122
|
+
sigmas: Per-prototype bandwidths, shape (n_prototypes,).
|
|
123
|
+
|
|
124
|
+
Returns:
|
|
125
|
+
Squared distances in feature space, shape (n_samples, n_prototypes).
|
|
126
|
+
|
|
127
|
+
References
|
|
128
|
+
----------
|
|
129
|
+
.. [1] Villmann, T., Haase, S., & Kaden, M. (2015). Kernelized vector
|
|
130
|
+
quantization in gradient-descent learning. Neurocomputing.
|
|
131
|
+
"""
|
|
132
|
+
diff = X[:, None, :] - W[None, :, :] # (n, p, d)
|
|
133
|
+
sq_norms = jnp.sum(diff ** 2, axis=2) # (n, p)
|
|
134
|
+
K = jnp.exp(-sq_norms / (2.0 * sigmas[None, :] ** 2)) # (n, p)
|
|
135
|
+
return 2.0 * (1.0 - K)
|
|
136
|
+
|
|
137
|
+
|
|
138
|
+
@jit
|
|
139
|
+
def kernel_distance_squared_relevance(
|
|
140
|
+
X: chex.Array, W: chex.Array, sigmas: chex.Array,
|
|
141
|
+
relevances: chex.Array
|
|
142
|
+
) -> chex.Array:
|
|
143
|
+
"""
|
|
144
|
+
Squared kernel distance with relevance weighting and per-prototype bandwidth.
|
|
145
|
+
|
|
146
|
+
.. math::
|
|
147
|
+
|
|
148
|
+
d_\\kappa^2(x, w_k) = 2\\left(1 - \\exp\\left(-\\frac{\\sum_j \\lambda_j (x_j - w_{kj})^2}{2\\sigma_k^2}\\right)\\right)
|
|
149
|
+
|
|
150
|
+
Args:
|
|
151
|
+
X: Data matrix, shape (n_samples, n_features).
|
|
152
|
+
W: Prototype matrix, shape (n_prototypes, n_features).
|
|
153
|
+
sigmas: Per-prototype bandwidths, shape (n_prototypes,).
|
|
154
|
+
relevances: Normalized relevance weights, shape (n_features,).
|
|
155
|
+
|
|
156
|
+
Returns:
|
|
157
|
+
Squared distances in feature space, shape (n_samples, n_prototypes).
|
|
158
|
+
|
|
159
|
+
References
|
|
160
|
+
----------
|
|
161
|
+
.. [1] Villmann, T., Haase, S., & Kaden, M. (2015). Kernelized vector
|
|
162
|
+
quantization in gradient-descent learning. Neurocomputing.
|
|
163
|
+
"""
|
|
164
|
+
diff = X[:, None, :] - W[None, :, :] # (n, p, d)
|
|
165
|
+
weighted_sq = jnp.sum(relevances[None, None, :] * diff ** 2, axis=2) # (n, p)
|
|
166
|
+
K = jnp.exp(-weighted_sq / (2.0 * sigmas[None, :] ** 2)) # (n, p)
|
|
167
|
+
return 2.0 * (1.0 - K)
|
|
168
|
+
|
|
169
|
+
|
|
170
|
+
@jit
|
|
171
|
+
def exponential_kernel_distance_squared(
|
|
172
|
+
X: chex.Array, W: chex.Array, omega_hat: chex.Array
|
|
173
|
+
) -> chex.Array:
|
|
174
|
+
"""
|
|
175
|
+
Squared distance in exponential kernel feature space.
|
|
176
|
+
|
|
177
|
+
Uses the exponential kernel :math:`\\kappa_{\\exp}(v, w, \\hat\\Lambda) = \\exp(v^T \\hat\\Lambda w)`
|
|
178
|
+
where :math:`\\hat\\Lambda = \\hat\\Omega \\hat\\Omega^T`.
|
|
179
|
+
|
|
180
|
+
.. math::
|
|
181
|
+
|
|
182
|
+
d_\\kappa^2(x, w) = \\exp(x^T \\hat\\Lambda x)
|
|
183
|
+
+ \\exp(w^T \\hat\\Lambda w)
|
|
184
|
+
- 2 \\exp(x^T \\hat\\Lambda w)
|
|
185
|
+
|
|
186
|
+
Note: :math:`\\kappa(v, v) \\neq 1` for the exponential kernel, so the
|
|
187
|
+
full three-term formula is required (not the 2(1-K) simplification).
|
|
188
|
+
|
|
189
|
+
Args:
|
|
190
|
+
X: Data matrix, shape (n_samples, n_features).
|
|
191
|
+
W: Prototype matrix, shape (n_prototypes, n_features).
|
|
192
|
+
omega_hat: Transformation matrix, shape (n_features, latent_dim).
|
|
193
|
+
The kernel matrix is :math:`\\hat\\Lambda = \\hat\\Omega \\hat\\Omega^T`.
|
|
194
|
+
|
|
195
|
+
Returns:
|
|
196
|
+
Squared distances in feature space, shape (n_samples, n_prototypes).
|
|
197
|
+
|
|
198
|
+
References
|
|
199
|
+
----------
|
|
200
|
+
.. [1] Villmann, T., Haase, S., & Kaden, M. (2015). Kernelized vector
|
|
201
|
+
quantization in gradient-descent learning. Neurocomputing.
|
|
202
|
+
"""
|
|
203
|
+
# Λ̂ = Ω̂ Ω̂^T (d, d)
|
|
204
|
+
lambda_hat = jnp.dot(omega_hat, omega_hat.T)
|
|
205
|
+
|
|
206
|
+
# x^T Λ̂ x for all samples: (n,)
|
|
207
|
+
Lx = jnp.dot(X, lambda_hat) # (n, d)
|
|
208
|
+
xLx = jnp.sum(X * Lx, axis=1) # (n,)
|
|
209
|
+
|
|
210
|
+
# w^T Λ̂ w for all prototypes: (p,)
|
|
211
|
+
Lw = jnp.dot(W, lambda_hat) # (p, d)
|
|
212
|
+
wLw = jnp.sum(W * Lw, axis=1) # (p,)
|
|
213
|
+
|
|
214
|
+
# x^T Λ̂ w for all (n, p) pairs: (n, p)
|
|
215
|
+
xLw = jnp.dot(Lx, W.T) # (n, p)
|
|
216
|
+
|
|
217
|
+
# d_κ²(x, w) = exp(x^T Λ̂ x) + exp(w^T Λ̂ w) - 2·exp(x^T Λ̂ w)
|
|
218
|
+
distances = (jnp.exp(xLx[:, None])
|
|
219
|
+
+ jnp.exp(wLw[None, :])
|
|
220
|
+
- 2.0 * jnp.exp(xLw))
|
|
221
|
+
|
|
222
|
+
return jnp.maximum(distances, 0.0)
|
|
@@ -74,7 +74,7 @@ def glvq_loss(distances, target_labels, prototype_labels,
|
|
|
74
74
|
margin=0.0):
|
|
75
75
|
"""Generalized LVQ loss.
|
|
76
76
|
|
|
77
|
-
mu_i =
|
|
77
|
+
:math:`\mu_i = \frac{d^+_i - d^-_i}{d^+_i + d^-_i}`
|
|
78
78
|
|
|
79
79
|
Parameters
|
|
80
80
|
----------
|
|
@@ -98,7 +98,7 @@ def glvq_loss_with_transfer(distances, target_labels, prototype_labels,
|
|
|
98
98
|
transfer_fn=identity, margin=0.0, beta=10.0):
|
|
99
99
|
"""GLVQ loss with configurable transfer function.
|
|
100
100
|
|
|
101
|
-
loss = mean(
|
|
101
|
+
:math:`\text{loss} = \text{mean}(f(\mu + \text{margin}, \beta))`
|
|
102
102
|
|
|
103
103
|
Parameters
|
|
104
104
|
----------
|
|
@@ -198,7 +198,7 @@ def _class_probabilities(distances, target_labels, prototype_labels, sigma):
|
|
|
198
198
|
def nllr_loss(distances, target_labels, prototype_labels, sigma=1.0):
|
|
199
199
|
"""Negative Log-Likelihood Ratio loss (for SLVQ).
|
|
200
200
|
|
|
201
|
-
loss =
|
|
201
|
+
:math:`\text{loss} = -\log(P(\text{correct}) / P(\text{wrong}))`
|
|
202
202
|
|
|
203
203
|
Parameters
|
|
204
204
|
----------
|
|
@@ -223,7 +223,7 @@ def nllr_loss(distances, target_labels, prototype_labels, sigma=1.0):
|
|
|
223
223
|
def rslvq_loss(distances, target_labels, prototype_labels, sigma=1.0):
|
|
224
224
|
"""Robust Soft LVQ loss (for RSLVQ).
|
|
225
225
|
|
|
226
|
-
loss =
|
|
226
|
+
:math:`\text{loss} = -\log(P(\text{correct}) / P(\text{all}))`
|
|
227
227
|
|
|
228
228
|
Parameters
|
|
229
229
|
----------
|
|
@@ -336,7 +336,7 @@ def cross_entropy_lvq_loss(distances, target_labels, prototype_labels, n_classes
|
|
|
336
336
|
def margin_loss(y_pred, y_true_one_hot, margin=0.3):
|
|
337
337
|
"""Margin loss for CBC.
|
|
338
338
|
|
|
339
|
-
loss =
|
|
339
|
+
:math:`\text{loss} = \text{ReLU}(\max(\text{wrong}) - \text{correct} + \text{margin})`
|
|
340
340
|
|
|
341
341
|
Parameters
|
|
342
342
|
----------
|
|
@@ -13,7 +13,7 @@ from jax import jit
|
|
|
13
13
|
def gaussian_similarity(distances_sq, variance=1.0):
|
|
14
14
|
"""Convert squared distances to Gaussian similarities.
|
|
15
15
|
|
|
16
|
-
s(d) = exp(-d^2 / (2
|
|
16
|
+
:math:`s(d) = \exp(-d^2 / (2 \cdot \text{variance}))`
|
|
17
17
|
|
|
18
18
|
Parameters
|
|
19
19
|
----------
|
|
@@ -34,7 +34,7 @@ def gaussian_similarity(distances_sq, variance=1.0):
|
|
|
34
34
|
def cosine_similarity_matrix(X, Y):
|
|
35
35
|
"""Pairwise cosine similarity between rows of X and Y.
|
|
36
36
|
|
|
37
|
-
cos(x, y) =
|
|
37
|
+
:math:`\cos(x, y) = \frac{x \cdot y}{\|x\| \cdot \|y\|}`
|
|
38
38
|
|
|
39
39
|
Parameters
|
|
40
40
|
----------
|
|
@@ -92,6 +92,14 @@ try:
|
|
|
92
92
|
from .heskes_som import HeskesSOM
|
|
93
93
|
from .riemannian_neural_gas import RiemannianNeuralGas
|
|
94
94
|
|
|
95
|
+
# Differentiating Kernel models (Villmann, Haase & Kaden, 2015)
|
|
96
|
+
from .dk_glvq import DKGLVQ
|
|
97
|
+
from .dk_relevance_lvq import DKGRLVQ
|
|
98
|
+
from .dk_matrix_lvq import DKGMLVQ
|
|
99
|
+
from .dk_neural_gas import DKNeuralGas
|
|
100
|
+
from .dk_kohonen_som import DKKohonenSOM
|
|
101
|
+
from .dk_heskes_som import DKHeskesSOM
|
|
102
|
+
|
|
95
103
|
_MODEL_REGISTRY = {
|
|
96
104
|
# Unsupervised clustering
|
|
97
105
|
'FCM': FCM, 'PCM': PCM, 'PFCM': PFCM, 'AFCM': AFCM,
|
|
@@ -134,6 +142,10 @@ try:
|
|
|
134
142
|
'NeuralGas': NeuralGas, 'GrowingNeuralGas': GrowingNeuralGas,
|
|
135
143
|
'KohonenSOM': KohonenSOM, 'HeskesSOM': HeskesSOM,
|
|
136
144
|
'RiemannianNeuralGas': RiemannianNeuralGas,
|
|
145
|
+
# Differentiating Kernel
|
|
146
|
+
'DKGLVQ': DKGLVQ, 'DKGRLVQ': DKGRLVQ, 'DKGMLVQ': DKGMLVQ,
|
|
147
|
+
'DKNeuralGas': DKNeuralGas, 'DKKohonenSOM': DKKohonenSOM,
|
|
148
|
+
'DKHeskesSOM': DKHeskesSOM,
|
|
137
149
|
}
|
|
138
150
|
|
|
139
151
|
JAX_AVAILABLE = True
|
|
@@ -180,6 +192,9 @@ __all__ = [
|
|
|
180
192
|
# Unsupervised topology
|
|
181
193
|
'NeuralGas', 'GrowingNeuralGas', 'KohonenSOM', 'HeskesSOM',
|
|
182
194
|
'RiemannianNeuralGas',
|
|
195
|
+
# Differentiating Kernel
|
|
196
|
+
'DKGLVQ', 'DKGRLVQ', 'DKGMLVQ',
|
|
197
|
+
'DKNeuralGas', 'DKKohonenSOM', 'DKHeskesSOM',
|
|
183
198
|
# Status
|
|
184
199
|
'JAX_AVAILABLE',
|
|
185
200
|
]
|
|
@@ -0,0 +1,246 @@
|
|
|
1
|
+
"""
|
|
2
|
+
Differentiating Kernel GLVQ (DKGLVQ).
|
|
3
|
+
|
|
4
|
+
GLVQ with Gaussian kernel distance and per-prototype bandwidth adaptation.
|
|
5
|
+
Each prototype :math:`w_k` has a learnable bandwidth :math:`\\sigma_k`:
|
|
6
|
+
|
|
7
|
+
.. math::
|
|
8
|
+
|
|
9
|
+
d_\\kappa^2(x, w_k) = 2\\left(1 - \\exp\\left(
|
|
10
|
+
-\\frac{\\|x - w_k\\|^2}{2\\sigma_k^2}
|
|
11
|
+
\\right)\\right)
|
|
12
|
+
|
|
13
|
+
References
|
|
14
|
+
----------
|
|
15
|
+
.. [1] Villmann, T., Haase, S., & Kaden, M. (2015). Kernelized vector
|
|
16
|
+
quantization in gradient-descent learning. Neurocomputing.
|
|
17
|
+
"""
|
|
18
|
+
|
|
19
|
+
import jax
|
|
20
|
+
import jax.numpy as jnp
|
|
21
|
+
from jax import jit
|
|
22
|
+
import numpy as np
|
|
23
|
+
|
|
24
|
+
from prosemble.models.prototype_base import SupervisedPrototypeModel
|
|
25
|
+
from prosemble.core.competitions import wtac
|
|
26
|
+
from prosemble.core.kernel import kernel_distance_squared_per_proto
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
@jit
|
|
30
|
+
def _predict_dkglvq_jit(X, prototypes, sigmas, proto_labels, sigma_min):
|
|
31
|
+
"""JIT-compiled DKGLVQ prediction with per-prototype kernel distance."""
|
|
32
|
+
sigmas = jnp.maximum(sigmas, sigma_min)
|
|
33
|
+
distances = kernel_distance_squared_per_proto(X, prototypes, sigmas)
|
|
34
|
+
return wtac(distances, proto_labels)
|
|
35
|
+
|
|
36
|
+
|
|
37
|
+
class DKGLVQ(SupervisedPrototypeModel):
|
|
38
|
+
"""Differentiating Kernel GLVQ.
|
|
39
|
+
|
|
40
|
+
GLVQ with Gaussian kernel distance in feature space. Each prototype
|
|
41
|
+
:math:`w_k` has a learnable bandwidth :math:`\\sigma_k` that is
|
|
42
|
+
adapted via gradient descent alongside the prototypes.
|
|
43
|
+
|
|
44
|
+
.. math::
|
|
45
|
+
|
|
46
|
+
d_\\kappa^2(x, w_k) = 2\\left(1 - \\exp\\left(
|
|
47
|
+
-\\frac{\\|x - w_k\\|^2}{2\\sigma_k^2}
|
|
48
|
+
\\right)\\right)
|
|
49
|
+
|
|
50
|
+
Parameters
|
|
51
|
+
----------
|
|
52
|
+
sigma_init : str or float
|
|
53
|
+
Initialization strategy for per-prototype bandwidths.
|
|
54
|
+
'median' (default): per-class median distance from prototype to class members.
|
|
55
|
+
'mean': per-class mean distance.
|
|
56
|
+
float: fixed value for all prototypes.
|
|
57
|
+
sigma_min : float
|
|
58
|
+
Lower bound for sigma to prevent bandwidth collapse. Default: 1e-3.
|
|
59
|
+
beta : float
|
|
60
|
+
Transfer function steepness parameter.
|
|
61
|
+
n_prototypes_per_class : int
|
|
62
|
+
Number of prototypes per class.
|
|
63
|
+
max_iter : int
|
|
64
|
+
Maximum training iterations.
|
|
65
|
+
lr : float
|
|
66
|
+
Learning rate.
|
|
67
|
+
epsilon : float
|
|
68
|
+
Convergence threshold on loss change.
|
|
69
|
+
random_seed : int
|
|
70
|
+
Random seed for reproducibility.
|
|
71
|
+
optimizer : str or optax optimizer, optional
|
|
72
|
+
Optimizer name ('adam', 'sgd') or optax GradientTransformation.
|
|
73
|
+
transfer_fn : callable, optional
|
|
74
|
+
Transfer function for loss shaping (default: identity).
|
|
75
|
+
margin : float
|
|
76
|
+
Margin for loss computation.
|
|
77
|
+
callbacks : list, optional
|
|
78
|
+
List of Callback objects.
|
|
79
|
+
use_scan : bool
|
|
80
|
+
If True (default), use jax.lax.scan for training.
|
|
81
|
+
batch_size : int, optional
|
|
82
|
+
Mini-batch size. If None, use full-batch training.
|
|
83
|
+
lr_scheduler : str or optax.Schedule, optional
|
|
84
|
+
Learning rate schedule.
|
|
85
|
+
lr_scheduler_kwargs : dict, optional
|
|
86
|
+
Keyword arguments for the learning rate scheduler.
|
|
87
|
+
prototypes_initializer : str or callable, optional
|
|
88
|
+
How to initialize prototypes.
|
|
89
|
+
patience : int, optional
|
|
90
|
+
Epochs with no improvement before stopping.
|
|
91
|
+
restore_best : bool
|
|
92
|
+
If True, restore best parameters after training.
|
|
93
|
+
class_weight : dict or 'balanced', optional
|
|
94
|
+
Weights for each class.
|
|
95
|
+
gradient_accumulation_steps : int, optional
|
|
96
|
+
Accumulate gradients over this many steps.
|
|
97
|
+
ema_decay : float, optional
|
|
98
|
+
Exponential moving average decay for parameters.
|
|
99
|
+
freeze_params : list of str, optional
|
|
100
|
+
Parameter group names to freeze.
|
|
101
|
+
lookahead : dict, optional
|
|
102
|
+
Lookahead optimizer wrapper configuration.
|
|
103
|
+
mixed_precision : str or None, optional
|
|
104
|
+
Compute dtype for mixed precision training.
|
|
105
|
+
|
|
106
|
+
References
|
|
107
|
+
----------
|
|
108
|
+
.. [1] Villmann, T., Haase, S., & Kaden, M. (2015). Kernelized vector
|
|
109
|
+
quantization in gradient-descent learning. Neurocomputing.
|
|
110
|
+
|
|
111
|
+
See Also
|
|
112
|
+
--------
|
|
113
|
+
SupervisedPrototypeModel : Full list of base parameters.
|
|
114
|
+
"""
|
|
115
|
+
|
|
116
|
+
def __init__(self, sigma_init='median', sigma_min=1e-3, beta=10.0,
|
|
117
|
+
n_prototypes_per_class=1, max_iter=100,
|
|
118
|
+
lr=0.01, epsilon=1e-6, random_seed=42, distance_fn=None,
|
|
119
|
+
optimizer='adam', transfer_fn=None, margin=0.0,
|
|
120
|
+
callbacks=None, use_scan=True, batch_size=None,
|
|
121
|
+
lr_scheduler=None, lr_scheduler_kwargs=None,
|
|
122
|
+
prototypes_initializer=None, patience=None,
|
|
123
|
+
restore_best=False, class_weight=None,
|
|
124
|
+
gradient_accumulation_steps=None, ema_decay=None,
|
|
125
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
126
|
+
**kwargs):
|
|
127
|
+
super().__init__(
|
|
128
|
+
n_prototypes_per_class=n_prototypes_per_class,
|
|
129
|
+
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
130
|
+
random_seed=random_seed, distance_fn=distance_fn,
|
|
131
|
+
optimizer=optimizer, transfer_fn=transfer_fn, margin=margin,
|
|
132
|
+
callbacks=callbacks, use_scan=use_scan, batch_size=batch_size,
|
|
133
|
+
lr_scheduler=lr_scheduler,
|
|
134
|
+
lr_scheduler_kwargs=lr_scheduler_kwargs,
|
|
135
|
+
prototypes_initializer=prototypes_initializer,
|
|
136
|
+
patience=patience, restore_best=restore_best,
|
|
137
|
+
class_weight=class_weight,
|
|
138
|
+
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
139
|
+
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
140
|
+
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
141
|
+
**kwargs,
|
|
142
|
+
)
|
|
143
|
+
self.sigma_init = sigma_init
|
|
144
|
+
self.sigma_min = sigma_min
|
|
145
|
+
self.beta = beta
|
|
146
|
+
self.sigmas_ = None
|
|
147
|
+
|
|
148
|
+
def _estimate_sigmas(self, X, y, prototypes, proto_labels):
|
|
149
|
+
"""Estimate per-prototype bandwidths from data."""
|
|
150
|
+
if isinstance(self.sigma_init, (int, float)):
|
|
151
|
+
return jnp.full(prototypes.shape[0], float(self.sigma_init))
|
|
152
|
+
|
|
153
|
+
sigmas = []
|
|
154
|
+
for k in range(prototypes.shape[0]):
|
|
155
|
+
label_k = proto_labels[k]
|
|
156
|
+
class_mask = (y == label_k)
|
|
157
|
+
X_class = X[class_mask]
|
|
158
|
+
dists = jnp.sqrt(jnp.sum((X_class - prototypes[k]) ** 2, axis=1))
|
|
159
|
+
if self.sigma_init == 'median':
|
|
160
|
+
sigma_k = jnp.median(dists)
|
|
161
|
+
else: # 'mean'
|
|
162
|
+
sigma_k = jnp.mean(dists)
|
|
163
|
+
sigmas.append(jnp.maximum(sigma_k, self.sigma_min))
|
|
164
|
+
return jnp.array(sigmas)
|
|
165
|
+
|
|
166
|
+
def _get_resume_params(self, params):
|
|
167
|
+
params['sigmas'] = self.sigmas_
|
|
168
|
+
return params
|
|
169
|
+
|
|
170
|
+
def _init_state(self, X, y, key):
|
|
171
|
+
key1, key2 = jax.random.split(key)
|
|
172
|
+
prototypes, proto_labels = self._init_prototypes(
|
|
173
|
+
X, y, self.n_prototypes_per_class, key1
|
|
174
|
+
)
|
|
175
|
+
sigmas = self._estimate_sigmas(X, y, prototypes, proto_labels)
|
|
176
|
+
params = {'prototypes': prototypes, 'sigmas': sigmas}
|
|
177
|
+
opt_state = self._optimizer.init(params)
|
|
178
|
+
from prosemble.models.prototype_base import SupervisedState
|
|
179
|
+
state = SupervisedState(
|
|
180
|
+
prototypes=prototypes,
|
|
181
|
+
opt_state=opt_state,
|
|
182
|
+
loss=jnp.array(float('inf')),
|
|
183
|
+
iteration=0,
|
|
184
|
+
converged=False,
|
|
185
|
+
)
|
|
186
|
+
return state, params, proto_labels
|
|
187
|
+
|
|
188
|
+
def _compute_loss(self, params, X, y, proto_labels):
|
|
189
|
+
prototypes = params['prototypes']
|
|
190
|
+
sigmas = jnp.maximum(params['sigmas'], self.sigma_min)
|
|
191
|
+
distances = kernel_distance_squared_per_proto(X, prototypes, sigmas)
|
|
192
|
+
from prosemble.core.losses import glvq_loss_with_transfer
|
|
193
|
+
return glvq_loss_with_transfer(
|
|
194
|
+
distances, y, proto_labels,
|
|
195
|
+
transfer_fn=self.transfer_fn,
|
|
196
|
+
margin=self.margin,
|
|
197
|
+
beta=self.beta,
|
|
198
|
+
)
|
|
199
|
+
|
|
200
|
+
def _post_update(self, params):
|
|
201
|
+
params['sigmas'] = jnp.maximum(params['sigmas'], self.sigma_min)
|
|
202
|
+
return params
|
|
203
|
+
|
|
204
|
+
def _extract_results(self, params, proto_labels, loss_history, n_iter, **kwargs):
|
|
205
|
+
super()._extract_results(params, proto_labels, loss_history, n_iter, **kwargs)
|
|
206
|
+
self.sigmas_ = params['sigmas']
|
|
207
|
+
|
|
208
|
+
@property
|
|
209
|
+
def kernel_bandwidths(self):
|
|
210
|
+
"""Return the learned per-prototype bandwidths."""
|
|
211
|
+
if self.sigmas_ is None:
|
|
212
|
+
raise ValueError("Model not fitted. Call fit() first.")
|
|
213
|
+
return self.sigmas_
|
|
214
|
+
|
|
215
|
+
def predict(self, X):
|
|
216
|
+
"""Predict using learned kernel distance."""
|
|
217
|
+
self._check_fitted()
|
|
218
|
+
X = jnp.asarray(X, dtype=jnp.float32)
|
|
219
|
+
return _predict_dkglvq_jit(
|
|
220
|
+
X, self.prototypes_, self.sigmas_,
|
|
221
|
+
self.prototype_labels_, self.sigma_min,
|
|
222
|
+
)
|
|
223
|
+
|
|
224
|
+
def _get_quantizable_attrs(self):
|
|
225
|
+
attrs = super()._get_quantizable_attrs()
|
|
226
|
+
if self.sigmas_ is not None:
|
|
227
|
+
attrs.append('sigmas_')
|
|
228
|
+
return attrs
|
|
229
|
+
|
|
230
|
+
def _get_fitted_arrays(self):
|
|
231
|
+
arrays = super()._get_fitted_arrays()
|
|
232
|
+
if self.sigmas_ is not None:
|
|
233
|
+
arrays['sigmas_'] = np.asarray(self.sigmas_)
|
|
234
|
+
return arrays
|
|
235
|
+
|
|
236
|
+
def _set_fitted_arrays(self, arrays):
|
|
237
|
+
super()._set_fitted_arrays(arrays)
|
|
238
|
+
if 'sigmas_' in arrays:
|
|
239
|
+
self.sigmas_ = jnp.asarray(arrays['sigmas_'])
|
|
240
|
+
|
|
241
|
+
def _get_hyperparams(self):
|
|
242
|
+
hp = super()._get_hyperparams()
|
|
243
|
+
hp['beta'] = self.beta
|
|
244
|
+
hp['sigma_init'] = self.sigma_init
|
|
245
|
+
hp['sigma_min'] = self.sigma_min
|
|
246
|
+
return hp
|