prosemble 2.3.0__tar.gz → 2.4.0__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-2.3.0 → prosemble-2.4.0}/PKG-INFO +1 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/distance.py +183 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/manifolds.py +166 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/__init__.py +14 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/celvq_ng_mixin.py +11 -3
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_glvq_ng.py +17 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_matrix_lvq_ng.py +17 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_relevance_lvq_ng.py +17 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/mrslvq_ng.py +18 -6
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_dk_glvq.py +8 -6
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_dk_matrix_lvq.py +5 -4
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_dk_relevance_lvq.py +5 -4
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_glvq.py +56 -9
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_glvq_ng_mixin.py +10 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_gmlvq.py +20 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_grlvq.py +15 -4
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_gtlvq.py +15 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_lgmlvq.py +22 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_mrslvq.py +9 -5
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_rslvq.py +1 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_rslvq_ng_mixin.py +10 -4
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_glvq.py +6 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_m_slng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_m_smng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_m_stng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_matrix_lvq.py +10 -3
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_r_slng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_r_smng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_r_stng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_relevance_lvq.py +10 -3
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_slng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_smng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_dk_stng.py +12 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_neural_gas.py +6 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_slng.py +10 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_smng.py +15 -3
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_srng.py +100 -5
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/riemannian_stng.py +11 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/rslvq_ng.py +9 -3
- prosemble-2.4.0/prosemble/models/scmng.py +289 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/slng.py +17 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/smng.py +17 -2
- prosemble-2.4.0/prosemble/models/sng.py +259 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/srng.py +18 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/stng.py +17 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/svq_occ.py +2 -2
- prosemble-2.4.0/prosemble/models/wasserstein_glvq.py +231 -0
- prosemble-2.4.0/prosemble/models/wasserstein_gmlvq.py +270 -0
- prosemble-2.4.0/prosemble/models/wasserstein_grlvq.py +259 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble.egg-info/PKG-INFO +1 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble.egg-info/SOURCES.txt +11 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/pyproject.toml +1 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/api/models.rst +8 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/conf.py +1 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/onnx.rst +4 -1
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/riemannian.rst +91 -2
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/supervised.rst +147 -2
- prosemble-2.4.0/tests/test_hyperbolic.py +292 -0
- prosemble-2.4.0/tests/test_ng_lr_ratio.py +153 -0
- prosemble-2.4.0/tests/test_scmng.py +258 -0
- prosemble-2.4.0/tests/test_sng.py +225 -0
- prosemble-2.4.0/tests/test_wasserstein.py +302 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/uv.lock +202 -208
- {prosemble-2.3.0 → prosemble-2.4.0}/.bumpversion.cfg +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.editorconfig +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.envrc +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.github/ISSUE_TEMPLATE/config.yml +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.github/ISSUE_TEMPLATE/feature_request.md +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.github/workflows/ci.yml +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.github/workflows/release.yml +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.gitignore +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/.readthedocs.yaml +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/CONSTITUTION.md +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/LICENSE +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/MANIFEST.in +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/ONNX_COVERAGE.md +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/README.md +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/TASK.md +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/benchmarks/benchmark_distance_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/benchmarks/benchmark_fcm_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/benchmarks/benchmark_pcm_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/devenv.lock +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/devenv.nix +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/devenv.yaml +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/afcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/bgpc_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/cbc_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/celvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/fcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/fpcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/glvq_advanced_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/glvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/glvq_unequal_prototypes_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/gmlvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/grlvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/hcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/heskes_som_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/ipcm2_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/ipcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kafcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kfcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kfpcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kipcm2_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kipcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kmeans_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/knn_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kohonen_som_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kpcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/kpfcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/large_dataset_training.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/lcelvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/lmrslvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/lmrslvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/lvq1_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/lvqmln_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/mcelvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/mrslvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/mrslvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/neural_gas_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/npc_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/oc_lmrslvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/oc_lmrslvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/oc_mrslvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/oc_mrslvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/oc_rslvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/oc_rslvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/pcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/pfcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/plvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/rslvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/slng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/smng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/som_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/srng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/stng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/tcelvq_ng_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/vis_fcm_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/vis_glvq_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/vis_neural_gas_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/examples/vis_som_iris.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/flake.nix +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/mkdocs.yml +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/nix/package.nix +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/__init__.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/__init__.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/activations.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/callbacks.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/competitions.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/curriculum.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/data.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/distributed.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/geodesic.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/initializers.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/kernel.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/losses.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/model_selection.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/onnx_export.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/optimizers.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/pipeline.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/pooling.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/protocols.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/quantization.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/regularization.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/reject.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/serialization.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/similarities.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/utils.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/vis.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/core/visualizer.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/datasets/__init__.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/datasets/dataset.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/afcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/base.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/bgpc.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/cbc.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/celvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/crossentropy_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_glvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_heskes_som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_kohonen_som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_matrix_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_neural_gas.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/dk_relevance_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/fcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/fpcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/glvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/growing_neural_gas.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/hcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/heskes_som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/image_cbc.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/image_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/ipcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/ipcm2.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kafcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kfcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kfpcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kipcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kipcm2.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kmeans.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/knn.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kohonen_som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kpcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/kpfcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/lcelvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/local_matrix_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/lvq1.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/lvq21.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/lvqmln.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/matrix_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/matrix_rslvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/mcelvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/median_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/neural_gas.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/npc.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_dk_glvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_dk_matrix_lvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_dk_relevance_lvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_glvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_gmlvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_grlvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_gtlvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_lgmlvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_mrslvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/oc_rslvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/pcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/pfcm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/plvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/probabilistic_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/prototype_base.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/relevance_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/siamese_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/svq_occ_lm.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/svq_occ_m.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/svq_occ_r.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/svq_occ_t.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/tangent_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/models/tcelvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble/py.typed +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble.egg-info/dependency_links.txt +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble.egg-info/requires.txt +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/prosemble.egg-info/top_level.txt +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/requirements_docs.txt +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/setup.cfg +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/Makefile +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/api/core.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/advanced.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/clustering.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/differentiating_kernels.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/one_class.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/guides/topology.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/index.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/installation.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/make.bat +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/sphinx-docs/quickstart.rst +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/__init__.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_activations.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_cbc.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_celvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_checkpoint.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_clustering_patience.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_competitions.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_core.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_core_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_data.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_dataset.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_dataset_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_distance_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_distributed.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_dk_models.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_dk_ng_models.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_examples.sh +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_fcm_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_glvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_gmlvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_heskes_som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_image_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_improvements.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_initializers.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_jax_models.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_jit_inference.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_kohonen_som.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_lcelvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_losses.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_lvq_nongradient.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_lvqmln_plvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_matrix_rslvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_mcelvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_minibatch.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_mixed_precision.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_model_selection.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_mrslvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_neural_gas.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_dk_models.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_dk_ng_models.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_glvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_glvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_glvq_variants.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_mrslvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_rslvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_oc_rslvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_onnx_export.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_pcm_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_pfcm_jax.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_pipeline.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_pooling.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_probabilistic.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_protocols.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_prototype_base.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_quantization.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_riemannian_dk_models.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_riemannian_neural_gas.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_riemannian_supervised_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_rslvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_scan_training.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_serialization.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_siamese_lvq.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_similarities.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_slng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_smng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_srng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_stng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_svq_occ.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_svq_occ_r.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_svq_occ_variants.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_tcelvq_ng.py +0 -0
- {prosemble-2.3.0 → prosemble-2.4.0}/tests/test_vis.py +0 -0
|
@@ -431,6 +431,185 @@ def tangent_distance_matrix(
|
|
|
431
431
|
return jnp.sum(tangent_diff ** 2, axis=2)
|
|
432
432
|
|
|
433
433
|
|
|
434
|
+
# ============================================================================
|
|
435
|
+
# Wasserstein Distance Functions
|
|
436
|
+
# ============================================================================
|
|
437
|
+
|
|
438
|
+
|
|
439
|
+
@jit
|
|
440
|
+
def wasserstein2_distance_matrix(
|
|
441
|
+
X: chex.Array,
|
|
442
|
+
means: chex.Array,
|
|
443
|
+
log_variances: chex.Array
|
|
444
|
+
) -> chex.Array:
|
|
445
|
+
"""
|
|
446
|
+
Compute pairwise squared 2-Wasserstein distances from points to Gaussian prototypes.
|
|
447
|
+
|
|
448
|
+
Each prototype is a diagonal Gaussian :math:`\\mathcal{N}(\\mu_k, \\text{diag}(\\sigma_k^2))`.
|
|
449
|
+
Each input point :math:`x` is treated as a Dirac delta distribution :math:`\\delta_x`.
|
|
450
|
+
|
|
451
|
+
The squared 2-Wasserstein distance from a point to a diagonal Gaussian is:
|
|
452
|
+
|
|
453
|
+
.. math::
|
|
454
|
+
|
|
455
|
+
W_2^2(\\delta_x, \\mathcal{N}(\\mu_k, \\text{diag}(\\sigma_k^2)))
|
|
456
|
+
= \\sum_j (x_j - \\mu_{kj})^2 + \\sum_j \\sigma_{kj}^2
|
|
457
|
+
|
|
458
|
+
This decomposes into the squared Euclidean distance from the point to the
|
|
459
|
+
mean, plus the total variance (trace of covariance). Prototypes with
|
|
460
|
+
smaller variance are effectively "more certain" and attract nearby points
|
|
461
|
+
more strongly.
|
|
462
|
+
|
|
463
|
+
Parameters
|
|
464
|
+
----------
|
|
465
|
+
X : array of shape (n, d)
|
|
466
|
+
Data points.
|
|
467
|
+
means : array of shape (p, d)
|
|
468
|
+
Prototype mean vectors.
|
|
469
|
+
log_variances : array of shape (p, d)
|
|
470
|
+
Log of prototype variances (ensures positivity via ``exp``).
|
|
471
|
+
|
|
472
|
+
Returns
|
|
473
|
+
-------
|
|
474
|
+
D : array of shape (n, p)
|
|
475
|
+
Squared 2-Wasserstein distances.
|
|
476
|
+
|
|
477
|
+
References
|
|
478
|
+
----------
|
|
479
|
+
.. [1] Villani, C. (2009). Optimal Transport: Old and New.
|
|
480
|
+
Springer. Chapter 2.
|
|
481
|
+
.. [2] Givens, C. R. & Shortt, R. M. (1984). A class of Wasserstein
|
|
482
|
+
metrics for probability distributions. Michigan Math. J., 31(2).
|
|
483
|
+
"""
|
|
484
|
+
chex.assert_rank(X, 2)
|
|
485
|
+
chex.assert_rank(means, 2)
|
|
486
|
+
chex.assert_rank(log_variances, 2)
|
|
487
|
+
chex.assert_equal(X.shape[1], means.shape[1])
|
|
488
|
+
chex.assert_equal(means.shape, log_variances.shape)
|
|
489
|
+
|
|
490
|
+
# Squared Euclidean from points to means
|
|
491
|
+
eucl = squared_euclidean_distance_matrix(X, means) # (n, p)
|
|
492
|
+
|
|
493
|
+
# Variance spread penalty per prototype
|
|
494
|
+
variances = jnp.exp(log_variances) # (p, d)
|
|
495
|
+
spread = jnp.sum(variances, axis=1) # (p,)
|
|
496
|
+
|
|
497
|
+
return eucl + spread[None, :]
|
|
498
|
+
|
|
499
|
+
|
|
500
|
+
@jit
|
|
501
|
+
def wasserstein2_omega_distance_matrix(
|
|
502
|
+
X: chex.Array,
|
|
503
|
+
means: chex.Array,
|
|
504
|
+
log_variances: chex.Array,
|
|
505
|
+
omega: chex.Array
|
|
506
|
+
) -> chex.Array:
|
|
507
|
+
"""
|
|
508
|
+
Squared 2-Wasserstein distance with global metric adaptation.
|
|
509
|
+
|
|
510
|
+
Projects data and means through :math:`\\Omega` before computing
|
|
511
|
+
the Euclidean component, while variances contribute directly:
|
|
512
|
+
|
|
513
|
+
.. math::
|
|
514
|
+
|
|
515
|
+
W_2^2(x, k) = \\|\\Omega(x - \\mu_k)\\|^2 + \\sum_j \\sigma_{kj}^2
|
|
516
|
+
|
|
517
|
+
Parameters
|
|
518
|
+
----------
|
|
519
|
+
X : array of shape (n, d)
|
|
520
|
+
Data points.
|
|
521
|
+
means : array of shape (p, d)
|
|
522
|
+
Prototype mean vectors.
|
|
523
|
+
log_variances : array of shape (p, d)
|
|
524
|
+
Log of prototype variances.
|
|
525
|
+
omega : array of shape (d, l)
|
|
526
|
+
Global projection matrix.
|
|
527
|
+
|
|
528
|
+
Returns
|
|
529
|
+
-------
|
|
530
|
+
D : array of shape (n, p)
|
|
531
|
+
Squared 2-Wasserstein distances in projected space.
|
|
532
|
+
"""
|
|
533
|
+
chex.assert_rank(X, 2)
|
|
534
|
+
chex.assert_rank(means, 2)
|
|
535
|
+
chex.assert_rank(log_variances, 2)
|
|
536
|
+
chex.assert_rank(omega, 2)
|
|
537
|
+
chex.assert_equal(X.shape[1], means.shape[1])
|
|
538
|
+
chex.assert_equal(means.shape, log_variances.shape)
|
|
539
|
+
chex.assert_equal(X.shape[1], omega.shape[0])
|
|
540
|
+
|
|
541
|
+
# Project and compute squared Euclidean in projected space
|
|
542
|
+
X_proj = X @ omega # (n, l)
|
|
543
|
+
M_proj = means @ omega # (p, l)
|
|
544
|
+
eucl = squared_euclidean_distance_matrix(X_proj, M_proj) # (n, p)
|
|
545
|
+
|
|
546
|
+
# Variance spread penalty
|
|
547
|
+
variances = jnp.exp(log_variances) # (p, d)
|
|
548
|
+
spread = jnp.sum(variances, axis=1) # (p,)
|
|
549
|
+
|
|
550
|
+
return eucl + spread[None, :]
|
|
551
|
+
|
|
552
|
+
|
|
553
|
+
@jit
|
|
554
|
+
def wasserstein2_relevance_distance_matrix(
|
|
555
|
+
X: chex.Array,
|
|
556
|
+
means: chex.Array,
|
|
557
|
+
log_variances: chex.Array,
|
|
558
|
+
relevances: chex.Array
|
|
559
|
+
) -> chex.Array:
|
|
560
|
+
"""
|
|
561
|
+
Squared 2-Wasserstein distance with feature relevance weighting.
|
|
562
|
+
|
|
563
|
+
Applies per-feature relevance weights :math:`\\lambda_j` to the
|
|
564
|
+
Euclidean component:
|
|
565
|
+
|
|
566
|
+
.. math::
|
|
567
|
+
|
|
568
|
+
W_2^2(x, k) = \\sum_j \\lambda_j (x_j - \\mu_{kj})^2 + \\sum_j \\sigma_{kj}^2
|
|
569
|
+
|
|
570
|
+
where :math:`\\lambda_j = \\text{softmax}(r)_j` ensures non-negative
|
|
571
|
+
weights that sum to 1.
|
|
572
|
+
|
|
573
|
+
Parameters
|
|
574
|
+
----------
|
|
575
|
+
X : array of shape (n, d)
|
|
576
|
+
Data points.
|
|
577
|
+
means : array of shape (p, d)
|
|
578
|
+
Prototype mean vectors.
|
|
579
|
+
log_variances : array of shape (p, d)
|
|
580
|
+
Log of prototype variances.
|
|
581
|
+
relevances : array of shape (d,)
|
|
582
|
+
Raw relevance logits (softmax applied internally).
|
|
583
|
+
|
|
584
|
+
Returns
|
|
585
|
+
-------
|
|
586
|
+
D : array of shape (n, p)
|
|
587
|
+
Relevance-weighted squared 2-Wasserstein distances.
|
|
588
|
+
"""
|
|
589
|
+
chex.assert_rank(X, 2)
|
|
590
|
+
chex.assert_rank(means, 2)
|
|
591
|
+
chex.assert_rank(log_variances, 2)
|
|
592
|
+
chex.assert_rank(relevances, 1)
|
|
593
|
+
chex.assert_equal(X.shape[1], means.shape[1])
|
|
594
|
+
chex.assert_equal(means.shape, log_variances.shape)
|
|
595
|
+
chex.assert_equal(X.shape[1], relevances.shape[0])
|
|
596
|
+
|
|
597
|
+
# Softmax relevance weights
|
|
598
|
+
lambdas = jax.nn.softmax(relevances) # (d,)
|
|
599
|
+
|
|
600
|
+
# Weighted squared differences
|
|
601
|
+
# (n, 1, d) - (1, p, d) -> (n, p, d)
|
|
602
|
+
diff_sq = (X[:, None, :] - means[None, :, :]) ** 2
|
|
603
|
+
weighted = diff_sq * lambdas[None, None, :] # (n, p, d)
|
|
604
|
+
eucl = jnp.sum(weighted, axis=2) # (n, p)
|
|
605
|
+
|
|
606
|
+
# Variance spread penalty
|
|
607
|
+
variances = jnp.exp(log_variances) # (p, d)
|
|
608
|
+
spread = jnp.sum(variances, axis=1) # (p,)
|
|
609
|
+
|
|
610
|
+
return eucl + spread[None, :]
|
|
611
|
+
|
|
612
|
+
|
|
434
613
|
# ============================================================================
|
|
435
614
|
# Kernel Functions
|
|
436
615
|
# ============================================================================
|
|
@@ -764,6 +943,10 @@ __all__ = [
|
|
|
764
943
|
'omega_distance_matrix',
|
|
765
944
|
'lomega_distance_matrix',
|
|
766
945
|
'tangent_distance_matrix',
|
|
946
|
+
# Wasserstein distance functions
|
|
947
|
+
'wasserstein2_distance_matrix',
|
|
948
|
+
'wasserstein2_omega_distance_matrix',
|
|
949
|
+
'wasserstein2_relevance_distance_matrix',
|
|
767
950
|
# Kernel functions
|
|
768
951
|
'gaussian_kernel_matrix',
|
|
769
952
|
'polynomial_kernel_matrix',
|
|
@@ -7,6 +7,7 @@ for manifolds commonly arising in machine learning:
|
|
|
7
7
|
- SO(n): Special orthogonal group (rotation matrices)
|
|
8
8
|
- SPD(n): Symmetric positive definite matrices
|
|
9
9
|
- Grassmannian Gr(n, k): k-dimensional subspaces of R^n
|
|
10
|
+
- HyperbolicPoincare(d): Poincare ball model of hyperbolic space
|
|
10
11
|
|
|
11
12
|
All operations are JIT-compilable and vmap-compatible.
|
|
12
13
|
|
|
@@ -15,6 +16,9 @@ References
|
|
|
15
16
|
Schwarz, Psenickova, Villmann, Röhrbein (2026).
|
|
16
17
|
Topology-Preserving Prototype Learning on Riemannian Manifolds.
|
|
17
18
|
ESANN 2026.
|
|
19
|
+
|
|
20
|
+
Ganea, O., Becigneul, G., & Hofmann, T. (2018). Hyperbolic Neural
|
|
21
|
+
Networks. NeurIPS 2018.
|
|
18
22
|
"""
|
|
19
23
|
|
|
20
24
|
from __future__ import annotations
|
|
@@ -323,3 +327,165 @@ class Grassmannian:
|
|
|
323
327
|
def injectivity_radius(self, Q):
|
|
324
328
|
"""Injectivity radius of Gr(n,k) is :math:`\\pi/2`."""
|
|
325
329
|
return jnp.pi / 2.0
|
|
330
|
+
|
|
331
|
+
|
|
332
|
+
# ---------------------------------------------------------------------------
|
|
333
|
+
# HyperbolicPoincare(d) — Poincaré Ball Model of Hyperbolic Space
|
|
334
|
+
# ---------------------------------------------------------------------------
|
|
335
|
+
|
|
336
|
+
def _mobius_add(x, y, eps=1e-7):
|
|
337
|
+
"""Möbius addition in the Poincaré ball.
|
|
338
|
+
|
|
339
|
+
.. math::
|
|
340
|
+
|
|
341
|
+
x \\oplus y = \\frac{(1 + 2\\langle x, y\\rangle + \\|y\\|^2) x
|
|
342
|
+
+ (1 - \\|x\\|^2) y}
|
|
343
|
+
{1 + 2\\langle x, y\\rangle + \\|x\\|^2 \\|y\\|^2}
|
|
344
|
+
|
|
345
|
+
Parameters
|
|
346
|
+
----------
|
|
347
|
+
x, y : arrays of shape (d,)
|
|
348
|
+
|
|
349
|
+
Returns
|
|
350
|
+
-------
|
|
351
|
+
result : array of shape (d,)
|
|
352
|
+
"""
|
|
353
|
+
x_sq = jnp.sum(x ** 2)
|
|
354
|
+
y_sq = jnp.sum(y ** 2)
|
|
355
|
+
xy = jnp.sum(x * y)
|
|
356
|
+
num = (1.0 + 2.0 * xy + y_sq) * x + (1.0 - x_sq) * y
|
|
357
|
+
denom = 1.0 + 2.0 * xy + x_sq * y_sq
|
|
358
|
+
return num / (denom + eps)
|
|
359
|
+
|
|
360
|
+
|
|
361
|
+
def _conformal_factor(x, eps=1e-7):
|
|
362
|
+
"""Conformal factor :math:`\\lambda_x = 2 / (1 - \\|x\\|^2)`.
|
|
363
|
+
|
|
364
|
+
Parameters
|
|
365
|
+
----------
|
|
366
|
+
x : array of shape (d,)
|
|
367
|
+
|
|
368
|
+
Returns
|
|
369
|
+
-------
|
|
370
|
+
scalar
|
|
371
|
+
"""
|
|
372
|
+
return 2.0 / (1.0 - jnp.sum(x ** 2) + eps)
|
|
373
|
+
|
|
374
|
+
|
|
375
|
+
class HyperbolicPoincare:
|
|
376
|
+
"""Poincaré ball model of hyperbolic space :math:`\\mathbb{H}^d`.
|
|
377
|
+
|
|
378
|
+
Points are vectors in :math:`\\mathbb{B}^d = \\{x \\in \\mathbb{R}^d : \\|x\\| < 1\\}`.
|
|
379
|
+
The Riemannian metric is :math:`g_x = \\lambda_x^2 g_E` where
|
|
380
|
+
:math:`\\lambda_x = 2/(1-\\|x\\|^2)` is the conformal factor.
|
|
381
|
+
|
|
382
|
+
This manifold is the natural geometry for hierarchical and
|
|
383
|
+
tree-structured data — it can embed trees with arbitrarily low
|
|
384
|
+
distortion (Sarkar 2011).
|
|
385
|
+
|
|
386
|
+
Parameters
|
|
387
|
+
----------
|
|
388
|
+
d : int
|
|
389
|
+
Dimension of the hyperbolic space.
|
|
390
|
+
eps : float
|
|
391
|
+
Numerical safety margin for boundary clamping.
|
|
392
|
+
"""
|
|
393
|
+
|
|
394
|
+
def __init__(self, d: int, eps: float = 1e-5):
|
|
395
|
+
self.d = d
|
|
396
|
+
self.eps = eps
|
|
397
|
+
self.point_shape = (d,)
|
|
398
|
+
|
|
399
|
+
def distance(self, x, y):
|
|
400
|
+
"""Geodesic distance in the Poincaré ball.
|
|
401
|
+
|
|
402
|
+
.. math::
|
|
403
|
+
|
|
404
|
+
d(x, y) = \\operatorname{arcosh}\\!\\left(1 + \\frac{2\\|x - y\\|^2}
|
|
405
|
+
{(1 - \\|x\\|^2)(1 - \\|y\\|^2)}\\right)
|
|
406
|
+
"""
|
|
407
|
+
diff_sq = jnp.sum((x - y) ** 2)
|
|
408
|
+
x_sq = jnp.sum(x ** 2)
|
|
409
|
+
y_sq = jnp.sum(y ** 2)
|
|
410
|
+
denom = (1.0 - x_sq) * (1.0 - y_sq)
|
|
411
|
+
arg = 1.0 + 2.0 * diff_sq / (denom + self.eps)
|
|
412
|
+
# Stable arcosh with safe gradient: arcosh(z) = log(z + sqrt(z^2-1))
|
|
413
|
+
# Clamp arg >= 1 + eps to prevent gradient blow-up near z=1
|
|
414
|
+
arg = jnp.maximum(arg, 1.0 + 1e-6)
|
|
415
|
+
return jnp.arccosh(arg)
|
|
416
|
+
|
|
417
|
+
def distance_squared(self, x, y):
|
|
418
|
+
"""Squared geodesic distance."""
|
|
419
|
+
d = self.distance(x, y)
|
|
420
|
+
return d ** 2
|
|
421
|
+
|
|
422
|
+
def log_map(self, x, y):
|
|
423
|
+
"""Logarithmic map: tangent vector at x pointing toward y.
|
|
424
|
+
|
|
425
|
+
.. math::
|
|
426
|
+
|
|
427
|
+
\\text{Log}_x(y) = \\frac{2}{\\lambda_x}
|
|
428
|
+
\\operatorname{arctanh}(\\|-x \\oplus y\\|)
|
|
429
|
+
\\frac{-x \\oplus y}{\\|-x \\oplus y\\|}
|
|
430
|
+
"""
|
|
431
|
+
neg_x = -x
|
|
432
|
+
add_result = _mobius_add(neg_x, y, eps=self.eps)
|
|
433
|
+
norm = jnp.linalg.norm(add_result)
|
|
434
|
+
norm_safe = jnp.maximum(norm, 1e-10)
|
|
435
|
+
# arctanh is only defined for |z| < 1, clamp
|
|
436
|
+
norm_clamped = jnp.minimum(norm, 1.0 - 1e-7)
|
|
437
|
+
lam_x = _conformal_factor(x, eps=self.eps)
|
|
438
|
+
scale = (2.0 / lam_x) * jnp.arctanh(norm_clamped)
|
|
439
|
+
return scale * add_result / norm_safe
|
|
440
|
+
|
|
441
|
+
def exp_map(self, x, v):
|
|
442
|
+
"""Exponential map: move from x along tangent vector v.
|
|
443
|
+
|
|
444
|
+
.. math::
|
|
445
|
+
|
|
446
|
+
\\text{Exp}_x(v) = x \\oplus \\left(
|
|
447
|
+
\\tanh\\!\\left(\\frac{\\lambda_x \\|v\\|}{2}\\right)
|
|
448
|
+
\\frac{v}{\\|v\\|}\\right)
|
|
449
|
+
"""
|
|
450
|
+
v_norm = jnp.linalg.norm(v)
|
|
451
|
+
v_norm_safe = jnp.maximum(v_norm, 1e-10)
|
|
452
|
+
lam_x = _conformal_factor(x, eps=self.eps)
|
|
453
|
+
direction = v / v_norm_safe
|
|
454
|
+
t = jnp.tanh(lam_x * v_norm / 2.0) * direction
|
|
455
|
+
return _mobius_add(x, t, eps=self.eps)
|
|
456
|
+
|
|
457
|
+
def random_point(self, key):
|
|
458
|
+
"""Generate a random point uniformly in the Poincaré ball.
|
|
459
|
+
|
|
460
|
+
Uses the radial transform: sample direction uniformly on S^{d-1},
|
|
461
|
+
then radius r ~ Beta(1, d) to get uniform distribution in B^d,
|
|
462
|
+
scaled to stay safely inside the ball.
|
|
463
|
+
"""
|
|
464
|
+
key1, key2 = jax.random.split(key)
|
|
465
|
+
# Random direction on unit sphere
|
|
466
|
+
direction = jax.random.normal(key1, (self.d,))
|
|
467
|
+
direction = direction / (jnp.linalg.norm(direction) + 1e-10)
|
|
468
|
+
# Uniform radius in ball: r^d ~ U[0,1] => r = U^{1/d}
|
|
469
|
+
u = jax.random.uniform(key2, (), minval=0.01, maxval=0.95)
|
|
470
|
+
r = u ** (1.0 / self.d)
|
|
471
|
+
# Scale to stay inside ball with margin
|
|
472
|
+
r = r * (1.0 - self.eps)
|
|
473
|
+
return r * direction
|
|
474
|
+
|
|
475
|
+
def belongs(self, x):
|
|
476
|
+
"""Check if x is in the Poincaré ball: :math:`\\|x\\| < 1`."""
|
|
477
|
+
return jnp.sum(x ** 2) < 1.0
|
|
478
|
+
|
|
479
|
+
def project(self, x):
|
|
480
|
+
"""Project to the interior of the Poincaré ball.
|
|
481
|
+
|
|
482
|
+
Clamps the norm to be strictly less than 1.
|
|
483
|
+
"""
|
|
484
|
+
max_norm = 1.0 - self.eps
|
|
485
|
+
norm = jnp.linalg.norm(x)
|
|
486
|
+
scale = jnp.minimum(max_norm / (norm + 1e-10), 1.0)
|
|
487
|
+
return x * scale
|
|
488
|
+
|
|
489
|
+
def injectivity_radius(self, x):
|
|
490
|
+
"""Hyperbolic space has infinite injectivity radius."""
|
|
491
|
+
return jnp.inf
|
|
@@ -32,8 +32,10 @@ try:
|
|
|
32
32
|
# Supervised prototype models (LVQ family)
|
|
33
33
|
from .glvq import GLVQ, GLVQ1, GLVQ21
|
|
34
34
|
from .relevance_lvq import GRLVQ
|
|
35
|
+
from .sng import SNG
|
|
35
36
|
from .srng import SRNG
|
|
36
37
|
from .smng import SMNG
|
|
38
|
+
from .scmng import SCMNG
|
|
37
39
|
from .slng import SLNG
|
|
38
40
|
from .stng import STNG
|
|
39
41
|
from .matrix_lvq import GMLVQ
|
|
@@ -117,6 +119,12 @@ try:
|
|
|
117
119
|
from .dk_kohonen_som import DKKohonenSOM
|
|
118
120
|
from .dk_heskes_som import DKHeskesSOM
|
|
119
121
|
|
|
122
|
+
# Wasserstein LVQ models
|
|
123
|
+
from .wasserstein_glvq import WGLVQ
|
|
124
|
+
from .wasserstein_gmlvq import WGMLVQ
|
|
125
|
+
from .wasserstein_grlvq import WGRLVQ
|
|
126
|
+
|
|
127
|
+
|
|
120
128
|
# One-Class Differentiating Kernel models
|
|
121
129
|
from .oc_dk_glvq import OCDKGLVQ
|
|
122
130
|
from .oc_dk_relevance_lvq import OCDKGRLVQ
|
|
@@ -133,7 +141,7 @@ try:
|
|
|
133
141
|
'KPFCM': KPFCM, 'KIPCM': KIPCM, 'KIPCM2': KIPCM2,
|
|
134
142
|
# Supervised LVQ
|
|
135
143
|
'GLVQ': GLVQ, 'GLVQ1': GLVQ1, 'GLVQ21': GLVQ21,
|
|
136
|
-
'GRLVQ': GRLVQ, 'SRNG': SRNG, 'SMNG': SMNG, 'SLNG': SLNG, 'STNG': STNG,
|
|
144
|
+
'GRLVQ': GRLVQ, 'SNG': SNG, 'SRNG': SRNG, 'SMNG': SMNG, 'SCMNG': SCMNG, 'SLNG': SLNG, 'STNG': STNG,
|
|
137
145
|
'GMLVQ': GMLVQ, 'LGMLVQ': LGMLVQ,
|
|
138
146
|
'GTLVQ': GTLVQ, 'CELVQ': CELVQ, 'CELVQ_NG': CELVQ_NG,
|
|
139
147
|
'MCELVQ_NG': MCELVQ_NG, 'LCELVQ_NG': LCELVQ_NG, 'TCELVQ_NG': TCELVQ_NG,
|
|
@@ -191,6 +199,8 @@ try:
|
|
|
191
199
|
'OCDKGMLVQ': OCDKGMLVQ,
|
|
192
200
|
'OCDKGLVQ_NG': OCDKGLVQ_NG, 'OCDKGRLVQ_NG': OCDKGRLVQ_NG,
|
|
193
201
|
'OCDKGMLVQ_NG': OCDKGMLVQ_NG,
|
|
202
|
+
# Wasserstein LVQ
|
|
203
|
+
'WGLVQ': WGLVQ, 'WGMLVQ': WGMLVQ, 'WGRLVQ': WGRLVQ,
|
|
194
204
|
}
|
|
195
205
|
|
|
196
206
|
JAX_AVAILABLE = True
|
|
@@ -214,7 +224,7 @@ __all__ = [
|
|
|
214
224
|
'PCM', 'PFCM', 'SOM',
|
|
215
225
|
# Supervised LVQ family
|
|
216
226
|
'GLVQ', 'GLVQ1', 'GLVQ21',
|
|
217
|
-
'GRLVQ', 'SRNG', 'SMNG', 'SLNG', 'STNG',
|
|
227
|
+
'GRLVQ', 'SNG', 'SRNG', 'SMNG', 'SCMNG', 'SLNG', 'STNG',
|
|
218
228
|
'GMLVQ', 'LGMLVQ', 'GTLVQ',
|
|
219
229
|
'CELVQ', 'CELVQ_NG', 'MCELVQ_NG', 'LCELVQ_NG', 'TCELVQ_NG',
|
|
220
230
|
'LVQ1', 'LVQ21', 'MedianLVQ',
|
|
@@ -250,6 +260,8 @@ __all__ = [
|
|
|
250
260
|
# One-Class Differentiating Kernel
|
|
251
261
|
'OCDKGLVQ', 'OCDKGRLVQ', 'OCDKGMLVQ',
|
|
252
262
|
'OCDKGLVQ_NG', 'OCDKGRLVQ_NG', 'OCDKGMLVQ_NG',
|
|
263
|
+
# Wasserstein LVQ
|
|
264
|
+
'WGLVQ', 'WGMLVQ', 'WGRLVQ',
|
|
253
265
|
# Status
|
|
254
266
|
'JAX_AVAILABLE',
|
|
255
267
|
]
|
|
@@ -35,6 +35,9 @@ class CELVQNGMixin:
|
|
|
35
35
|
gamma_decay : float, optional
|
|
36
36
|
Per-step multiplicative decay for gamma.
|
|
37
37
|
Default: computed from max_iter so gamma reaches gamma_final.
|
|
38
|
+
lr_ratio : float
|
|
39
|
+
Ratio of wrong-class to correct-class learning rate (ε⁻/ε⁺).
|
|
40
|
+
Default: 0.5.
|
|
38
41
|
n_prototypes_per_class : int
|
|
39
42
|
Number of prototypes per class.
|
|
40
43
|
max_iter : int
|
|
@@ -157,7 +160,7 @@ class CELVQNGMixin:
|
|
|
157
160
|
params['gamma'] = jnp.array(gamma, dtype=jnp.float32)
|
|
158
161
|
return params
|
|
159
162
|
|
|
160
|
-
def _compute_gamma_init(self):
|
|
163
|
+
def _compute_gamma_init(self, X):
|
|
161
164
|
"""Compute gamma_init from prototype count if not set."""
|
|
162
165
|
if isinstance(self.n_prototypes_per_class, int):
|
|
163
166
|
max_per_class = self.n_prototypes_per_class
|
|
@@ -173,7 +176,12 @@ class CELVQNGMixin:
|
|
|
173
176
|
if self.gamma_decay is not None:
|
|
174
177
|
self._gamma_decay = self.gamma_decay
|
|
175
178
|
else:
|
|
176
|
-
|
|
179
|
+
if self.batch_size is not None:
|
|
180
|
+
steps_per_epoch = (X.shape[0] + self.batch_size - 1) // self.batch_size
|
|
181
|
+
else:
|
|
182
|
+
steps_per_epoch = 1
|
|
183
|
+
total_steps = self.max_iter * steps_per_epoch
|
|
184
|
+
self._gamma_decay = (self.gamma_final / gamma_init) ** (1.0 / total_steps)
|
|
177
185
|
|
|
178
186
|
return gamma_init
|
|
179
187
|
|
|
@@ -183,7 +191,7 @@ class CELVQNGMixin:
|
|
|
183
191
|
X, y, self.n_prototypes_per_class, key1
|
|
184
192
|
)
|
|
185
193
|
|
|
186
|
-
gamma_init = self._compute_gamma_init()
|
|
194
|
+
gamma_init = self._compute_gamma_init(X)
|
|
187
195
|
|
|
188
196
|
params = {
|
|
189
197
|
'prototypes': prototypes,
|
|
@@ -72,6 +72,9 @@ class DKGLVQ_NG(SupervisedPrototypeModel):
|
|
|
72
72
|
gamma_decay : float, optional
|
|
73
73
|
Per-step multiplicative decay for gamma.
|
|
74
74
|
Default: computed from max_iter so gamma reaches gamma_final.
|
|
75
|
+
lr_ratio : float
|
|
76
|
+
Ratio of wrong-class to correct-class learning rate (ε⁻/ε⁺).
|
|
77
|
+
Default: 0.5.
|
|
75
78
|
n_prototypes_per_class : int
|
|
76
79
|
Number of prototypes per class.
|
|
77
80
|
max_iter : int
|
|
@@ -140,7 +143,7 @@ class DKGLVQ_NG(SupervisedPrototypeModel):
|
|
|
140
143
|
|
|
141
144
|
def __init__(self, sigma_init='median', sigma_min=1e-3, beta=10.0,
|
|
142
145
|
gamma_init=None, gamma_final=0.01, gamma_decay=None,
|
|
143
|
-
n_prototypes_per_class=1, max_iter=100,
|
|
146
|
+
lr_ratio=0.5, n_prototypes_per_class=1, max_iter=100,
|
|
144
147
|
lr=0.01, epsilon=1e-6, random_seed=42, distance_fn=None,
|
|
145
148
|
optimizer='adam', transfer_fn=None, margin=0.0,
|
|
146
149
|
callbacks=None, use_scan=True, batch_size=None,
|
|
@@ -171,6 +174,7 @@ class DKGLVQ_NG(SupervisedPrototypeModel):
|
|
|
171
174
|
self.gamma_init = gamma_init
|
|
172
175
|
self.gamma_final = gamma_final
|
|
173
176
|
self.gamma_decay = gamma_decay
|
|
177
|
+
self.lr_ratio = lr_ratio
|
|
174
178
|
self.sigmas_ = None
|
|
175
179
|
self.gamma_ = None
|
|
176
180
|
|
|
@@ -227,7 +231,12 @@ class DKGLVQ_NG(SupervisedPrototypeModel):
|
|
|
227
231
|
if self.gamma_decay is not None:
|
|
228
232
|
self._gamma_decay = self.gamma_decay
|
|
229
233
|
else:
|
|
230
|
-
|
|
234
|
+
if self.batch_size is not None:
|
|
235
|
+
steps_per_epoch = (X.shape[0] + self.batch_size - 1) // self.batch_size
|
|
236
|
+
else:
|
|
237
|
+
steps_per_epoch = 1
|
|
238
|
+
total_steps = self.max_iter * steps_per_epoch
|
|
239
|
+
self._gamma_decay = (self.gamma_final / gamma_init) ** (1.0 / total_steps)
|
|
231
240
|
|
|
232
241
|
params = {
|
|
233
242
|
'prototypes': prototypes,
|
|
@@ -274,6 +283,11 @@ class DKGLVQ_NG(SupervisedPrototypeModel):
|
|
|
274
283
|
d_diff = jnp.where(~same_class, distances, INF)
|
|
275
284
|
dm = jnp.min(d_diff, axis=1)
|
|
276
285
|
|
|
286
|
+
# Separate learning rates (Hammer et al. 2003: ε⁻ = lr_ratio × ε⁺)
|
|
287
|
+
# Scale gradient through dm by lr_ratio; forward pass unchanged.
|
|
288
|
+
dm = jax.lax.stop_gradient(dm) + self.lr_ratio * (
|
|
289
|
+
dm - jax.lax.stop_gradient(dm))
|
|
290
|
+
|
|
277
291
|
# GLVQ mu for each (sample, prototype) pair
|
|
278
292
|
mu = (distances - dm[:, None]) / (distances + dm[:, None] + 1e-10)
|
|
279
293
|
|
|
@@ -353,4 +367,5 @@ class DKGLVQ_NG(SupervisedPrototypeModel):
|
|
|
353
367
|
hp['gamma_init'] = self.gamma_init
|
|
354
368
|
hp['gamma_final'] = self.gamma_final
|
|
355
369
|
hp['gamma_decay'] = self.gamma_decay
|
|
370
|
+
hp['lr_ratio'] = self.lr_ratio
|
|
356
371
|
return hp
|
|
@@ -68,6 +68,9 @@ class DKGMLVQ_NG(SupervisedPrototypeModel):
|
|
|
68
68
|
Final neighborhood range. Default: 0.01.
|
|
69
69
|
gamma_decay : float, optional
|
|
70
70
|
Per-step multiplicative decay for gamma.
|
|
71
|
+
lr_ratio : float
|
|
72
|
+
Ratio of wrong-class to correct-class learning rate (ε⁻/ε⁺).
|
|
73
|
+
Default: 0.5.
|
|
71
74
|
n_prototypes_per_class : int
|
|
72
75
|
Number of prototypes per class.
|
|
73
76
|
max_iter : int
|
|
@@ -136,7 +139,7 @@ class DKGMLVQ_NG(SupervisedPrototypeModel):
|
|
|
136
139
|
|
|
137
140
|
def __init__(self, latent_dim=None, omega_hat_scale=0.1, beta=10.0,
|
|
138
141
|
gamma_init=None, gamma_final=0.01, gamma_decay=None,
|
|
139
|
-
n_prototypes_per_class=1, max_iter=100,
|
|
142
|
+
lr_ratio=0.5, n_prototypes_per_class=1, max_iter=100,
|
|
140
143
|
lr=0.01, epsilon=1e-6, random_seed=42, distance_fn=None,
|
|
141
144
|
optimizer='adam', transfer_fn=None, margin=0.0,
|
|
142
145
|
callbacks=None, use_scan=True, batch_size=None,
|
|
@@ -167,6 +170,7 @@ class DKGMLVQ_NG(SupervisedPrototypeModel):
|
|
|
167
170
|
self.gamma_init = gamma_init
|
|
168
171
|
self.gamma_final = gamma_final
|
|
169
172
|
self.gamma_decay = gamma_decay
|
|
173
|
+
self.lr_ratio = lr_ratio
|
|
170
174
|
self.omega_hat_ = None
|
|
171
175
|
self.gamma_ = None
|
|
172
176
|
|
|
@@ -210,7 +214,12 @@ class DKGMLVQ_NG(SupervisedPrototypeModel):
|
|
|
210
214
|
if self.gamma_decay is not None:
|
|
211
215
|
self._gamma_decay = self.gamma_decay
|
|
212
216
|
else:
|
|
213
|
-
|
|
217
|
+
if self.batch_size is not None:
|
|
218
|
+
steps_per_epoch = (X.shape[0] + self.batch_size - 1) // self.batch_size
|
|
219
|
+
else:
|
|
220
|
+
steps_per_epoch = 1
|
|
221
|
+
total_steps = self.max_iter * steps_per_epoch
|
|
222
|
+
self._gamma_decay = (self.gamma_final / gamma_init) ** (1.0 / total_steps)
|
|
214
223
|
|
|
215
224
|
params = {
|
|
216
225
|
'prototypes': prototypes,
|
|
@@ -255,6 +264,11 @@ class DKGMLVQ_NG(SupervisedPrototypeModel):
|
|
|
255
264
|
d_diff = jnp.where(~same_class, distances, INF)
|
|
256
265
|
dm = jnp.min(d_diff, axis=1)
|
|
257
266
|
|
|
267
|
+
# Separate learning rates (Hammer et al. 2003: ε⁻ = lr_ratio × ε⁺)
|
|
268
|
+
# Scale gradient through dm by lr_ratio; forward pass unchanged.
|
|
269
|
+
dm = jax.lax.stop_gradient(dm) + self.lr_ratio * (
|
|
270
|
+
dm - jax.lax.stop_gradient(dm))
|
|
271
|
+
|
|
258
272
|
mu = (distances - dm[:, None]) / (distances + dm[:, None] + 1e-10)
|
|
259
273
|
|
|
260
274
|
from prosemble.core.activations import sigmoid_beta
|
|
@@ -331,6 +345,7 @@ class DKGMLVQ_NG(SupervisedPrototypeModel):
|
|
|
331
345
|
hp['gamma_init'] = self.gamma_init
|
|
332
346
|
hp['gamma_final'] = self.gamma_final
|
|
333
347
|
hp['gamma_decay'] = self.gamma_decay
|
|
348
|
+
hp['lr_ratio'] = self.lr_ratio
|
|
334
349
|
if self.latent_dim is not None:
|
|
335
350
|
hp['latent_dim'] = self.latent_dim
|
|
336
351
|
return hp
|
|
@@ -74,6 +74,9 @@ class DKGRLVQ_NG(SupervisedPrototypeModel):
|
|
|
74
74
|
Final neighborhood range. Default: 0.01.
|
|
75
75
|
gamma_decay : float, optional
|
|
76
76
|
Per-step multiplicative decay for gamma.
|
|
77
|
+
lr_ratio : float
|
|
78
|
+
Ratio of wrong-class to correct-class learning rate (ε⁻/ε⁺).
|
|
79
|
+
Default: 0.5.
|
|
77
80
|
n_prototypes_per_class : int
|
|
78
81
|
Number of prototypes per class.
|
|
79
82
|
max_iter : int
|
|
@@ -144,7 +147,7 @@ class DKGRLVQ_NG(SupervisedPrototypeModel):
|
|
|
144
147
|
|
|
145
148
|
def __init__(self, sigma_init='median', sigma_min=1e-3, beta=10.0,
|
|
146
149
|
gamma_init=None, gamma_final=0.01, gamma_decay=None,
|
|
147
|
-
n_prototypes_per_class=1, max_iter=100,
|
|
150
|
+
lr_ratio=0.5, n_prototypes_per_class=1, max_iter=100,
|
|
148
151
|
lr=0.01, epsilon=1e-6, random_seed=42, distance_fn=None,
|
|
149
152
|
optimizer='adam', transfer_fn=None, margin=0.0,
|
|
150
153
|
callbacks=None, use_scan=True, batch_size=None,
|
|
@@ -175,6 +178,7 @@ class DKGRLVQ_NG(SupervisedPrototypeModel):
|
|
|
175
178
|
self.gamma_init = gamma_init
|
|
176
179
|
self.gamma_final = gamma_final
|
|
177
180
|
self.gamma_decay = gamma_decay
|
|
181
|
+
self.lr_ratio = lr_ratio
|
|
178
182
|
self.sigmas_ = None
|
|
179
183
|
self.relevances_ = None
|
|
180
184
|
self.gamma_ = None
|
|
@@ -234,7 +238,12 @@ class DKGRLVQ_NG(SupervisedPrototypeModel):
|
|
|
234
238
|
if self.gamma_decay is not None:
|
|
235
239
|
self._gamma_decay = self.gamma_decay
|
|
236
240
|
else:
|
|
237
|
-
|
|
241
|
+
if self.batch_size is not None:
|
|
242
|
+
steps_per_epoch = (X.shape[0] + self.batch_size - 1) // self.batch_size
|
|
243
|
+
else:
|
|
244
|
+
steps_per_epoch = 1
|
|
245
|
+
total_steps = self.max_iter * steps_per_epoch
|
|
246
|
+
self._gamma_decay = (self.gamma_final / gamma_init) ** (1.0 / total_steps)
|
|
238
247
|
|
|
239
248
|
relevances = jnp.ones(n_features) / n_features
|
|
240
249
|
params = {
|
|
@@ -282,6 +291,11 @@ class DKGRLVQ_NG(SupervisedPrototypeModel):
|
|
|
282
291
|
d_diff = jnp.where(~same_class, distances, INF)
|
|
283
292
|
dm = jnp.min(d_diff, axis=1)
|
|
284
293
|
|
|
294
|
+
# Separate learning rates (Hammer et al. 2003: ε⁻ = lr_ratio × ε⁺)
|
|
295
|
+
# Scale gradient through dm by lr_ratio; forward pass unchanged.
|
|
296
|
+
dm = jax.lax.stop_gradient(dm) + self.lr_ratio * (
|
|
297
|
+
dm - jax.lax.stop_gradient(dm))
|
|
298
|
+
|
|
285
299
|
mu = (distances - dm[:, None]) / (distances + dm[:, None] + 1e-10)
|
|
286
300
|
|
|
287
301
|
from prosemble.core.activations import sigmoid_beta
|
|
@@ -373,4 +387,5 @@ class DKGRLVQ_NG(SupervisedPrototypeModel):
|
|
|
373
387
|
hp['gamma_init'] = self.gamma_init
|
|
374
388
|
hp['gamma_final'] = self.gamma_final
|
|
375
389
|
hp['gamma_decay'] = self.gamma_decay
|
|
390
|
+
hp['lr_ratio'] = self.lr_ratio
|
|
376
391
|
return hp
|