prosemble 2.2.0__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-2.2.0 → prosemble-2.4.0}/PKG-INFO +1 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/nix/package.nix +9 -8
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/__init__.py +25 -0
- prosemble-2.4.0/prosemble/core/curriculum.py +171 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/distance.py +183 -0
- prosemble-2.4.0/prosemble/core/geodesic.py +217 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/manifolds.py +166 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/onnx_export.py +146 -3
- prosemble-2.4.0/prosemble/core/optimizers.py +292 -0
- prosemble-2.4.0/prosemble/core/regularization.py +210 -0
- prosemble-2.4.0/prosemble/core/reject.py +246 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/__init__.py +47 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/celvq_ng_mixin.py +11 -3
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_glvq.py +5 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_glvq_ng.py +22 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_matrix_lvq.py +4 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_matrix_lvq_ng.py +21 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_relevance_lvq.py +6 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_relevance_lvq_ng.py +23 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/local_matrix_lvq.py +6 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/matrix_lvq.py +6 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/mrslvq_ng.py +18 -6
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_dk_glvq.py +8 -6
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_dk_matrix_lvq.py +5 -4
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_dk_relevance_lvq.py +5 -4
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_glvq.py +56 -9
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_glvq_ng_mixin.py +10 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_gmlvq.py +20 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_grlvq.py +15 -4
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_gtlvq.py +15 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_lgmlvq.py +22 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_mrslvq.py +9 -5
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_rslvq.py +1 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_rslvq_ng_mixin.py +10 -4
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/prototype_base.py +2 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/relevance_lvq.py +5 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_glvq.py +315 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_m_slng.py +314 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_m_smng.py +314 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_m_stng.py +318 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_matrix_lvq.py +350 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_r_slng.py +363 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_r_smng.py +364 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_r_stng.py +370 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_relevance_lvq.py +387 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_slng.py +330 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_smng.py +330 -0
- prosemble-2.4.0/prosemble/models/riemannian_dk_stng.py +339 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/riemannian_neural_gas.py +6 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/riemannian_slng.py +10 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/riemannian_smng.py +15 -3
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/riemannian_srng.py +100 -5
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/riemannian_stng.py +11 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/rslvq_ng.py +9 -3
- prosemble-2.4.0/prosemble/models/scmng.py +289 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/slng.py +17 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/smng.py +17 -2
- prosemble-2.4.0/prosemble/models/sng.py +259 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/srng.py +18 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/stng.py +17 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/svq_occ.py +2 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/tangent_lvq.py +8 -0
- 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-2.2.0 → prosemble-2.4.0}/prosemble.egg-info/PKG-INFO +1 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble.egg-info/SOURCES.txt +30 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/pyproject.toml +1 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/api/core.rst +35 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/api/models.rst +59 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/conf.py +1 -1
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/advanced.rst +262 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/differentiating_kernels.rst +74 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/onnx.rst +56 -2
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/riemannian.rst +91 -2
- {prosemble-2.2.0 → 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_improvements.py +703 -0
- prosemble-2.4.0/tests/test_ng_lr_ratio.py +153 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_onnx_export.py +90 -0
- prosemble-2.4.0/tests/test_riemannian_dk_models.py +1043 -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-2.2.0 → prosemble-2.4.0}/uv.lock +201 -207
- {prosemble-2.2.0 → prosemble-2.4.0}/.bumpversion.cfg +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.editorconfig +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.envrc +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.github/ISSUE_TEMPLATE/config.yml +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.github/ISSUE_TEMPLATE/feature_request.md +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.github/workflows/ci.yml +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.github/workflows/release.yml +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.gitignore +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/.readthedocs.yaml +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/CONSTITUTION.md +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/LICENSE +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/MANIFEST.in +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/ONNX_COVERAGE.md +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/README.md +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/TASK.md +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/benchmarks/benchmark_distance_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/benchmarks/benchmark_fcm_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/benchmarks/benchmark_pcm_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/devenv.lock +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/devenv.nix +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/devenv.yaml +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/afcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/bgpc_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/cbc_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/celvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/fcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/fpcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/glvq_advanced_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/glvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/glvq_unequal_prototypes_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/gmlvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/grlvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/hcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/heskes_som_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/ipcm2_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/ipcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kafcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kfcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kfpcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kipcm2_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kipcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kmeans_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/knn_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kohonen_som_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kpcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/kpfcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/large_dataset_training.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/lcelvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/lmrslvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/lmrslvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/lvq1_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/lvqmln_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/mcelvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/mrslvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/mrslvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/neural_gas_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/npc_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/oc_lmrslvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/oc_lmrslvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/oc_mrslvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/oc_mrslvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/oc_rslvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/oc_rslvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/pcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/pfcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/plvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/rslvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/slng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/smng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/som_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/srng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/stng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/tcelvq_ng_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/vis_fcm_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/vis_glvq_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/vis_neural_gas_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/examples/vis_som_iris.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/flake.nix +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/mkdocs.yml +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/__init__.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/activations.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/callbacks.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/competitions.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/data.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/distributed.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/initializers.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/kernel.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/losses.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/model_selection.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/pipeline.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/pooling.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/protocols.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/quantization.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/serialization.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/similarities.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/utils.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/vis.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/core/visualizer.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/datasets/__init__.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/datasets/dataset.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/afcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/base.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/bgpc.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/cbc.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/celvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/crossentropy_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_heskes_som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_kohonen_som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/dk_neural_gas.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/fcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/fpcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/glvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/growing_neural_gas.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/hcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/heskes_som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/image_cbc.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/image_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/ipcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/ipcm2.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kafcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kfcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kfpcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kipcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kipcm2.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kmeans.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/knn.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kohonen_som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kpcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/kpfcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/lcelvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/lvq1.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/lvq21.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/lvqmln.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/matrix_rslvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/mcelvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/median_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/neural_gas.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/npc.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_dk_glvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_dk_matrix_lvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_dk_relevance_lvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_glvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_gmlvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_grlvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_gtlvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_lgmlvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_mrslvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/oc_rslvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/pcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/pfcm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/plvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/probabilistic_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/siamese_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/svq_occ_lm.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/svq_occ_m.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/svq_occ_r.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/svq_occ_t.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/models/tcelvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble/py.typed +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble.egg-info/dependency_links.txt +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble.egg-info/requires.txt +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/prosemble.egg-info/top_level.txt +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/requirements_docs.txt +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/setup.cfg +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/Makefile +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/clustering.rst +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/one_class.rst +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/guides/topology.rst +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/index.rst +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/installation.rst +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/make.bat +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/sphinx-docs/quickstart.rst +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/__init__.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_activations.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_cbc.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_celvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_checkpoint.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_clustering_patience.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_competitions.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_core.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_core_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_data.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_dataset.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_dataset_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_distance_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_distributed.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_dk_models.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_dk_ng_models.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_examples.sh +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_fcm_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_glvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_gmlvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_heskes_som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_image_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_initializers.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_jax_models.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_jit_inference.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_kohonen_som.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_lcelvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_losses.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_lvq_nongradient.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_lvqmln_plvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_matrix_rslvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_mcelvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_minibatch.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_mixed_precision.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_model_selection.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_mrslvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_neural_gas.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_dk_models.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_dk_ng_models.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_glvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_glvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_glvq_variants.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_mrslvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_rslvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_oc_rslvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_pcm_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_pfcm_jax.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_pipeline.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_pooling.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_probabilistic.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_protocols.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_prototype_base.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_quantization.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_riemannian_neural_gas.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_riemannian_supervised_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_rslvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_scan_training.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_serialization.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_siamese_lvq.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_similarities.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_slng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_smng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_srng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_stng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_svq_occ.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_svq_occ_r.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_svq_occ_variants.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_tcelvq_ng.py +0 -0
- {prosemble-2.2.0 → prosemble-2.4.0}/tests/test_vis.py +0 -0
|
@@ -5,8 +5,8 @@
|
|
|
5
5
|
# 2. Add to pkgs/top-level/python-packages.nix:
|
|
6
6
|
# prosemble = callPackage ../development/python-modules/prosemble { };
|
|
7
7
|
# 3. Update the hash after publishing to PyPI:
|
|
8
|
-
# nix-prefetch-url --unpack https://github.com/naotoo1/prosemble/archive/refs/tags/v2.
|
|
9
|
-
# 4. PR title: python3Packages.prosemble: init at 2.
|
|
8
|
+
# nix-prefetch-url --unpack https://github.com/naotoo1/prosemble/archive/refs/tags/v2.3.0.tar.gz
|
|
9
|
+
# 4. PR title: python3Packages.prosemble: init at 2.3.0
|
|
10
10
|
{
|
|
11
11
|
lib,
|
|
12
12
|
buildPythonPackage,
|
|
@@ -29,14 +29,14 @@
|
|
|
29
29
|
|
|
30
30
|
buildPythonPackage (finalAttrs: {
|
|
31
31
|
pname = "prosemble";
|
|
32
|
-
version = "2.
|
|
32
|
+
version = "2.3.0";
|
|
33
33
|
pyproject = true;
|
|
34
34
|
|
|
35
35
|
src = fetchFromGitHub {
|
|
36
36
|
owner = "naotoo1";
|
|
37
37
|
repo = "prosemble";
|
|
38
|
-
|
|
39
|
-
hash = "sha256-
|
|
38
|
+
tag = "v${finalAttrs.version}";
|
|
39
|
+
hash = "sha256-296Z1lHBcOVQbkfO7KlNnWe7f7xjbKFHJRfGUPaHKWY=";
|
|
40
40
|
};
|
|
41
41
|
|
|
42
42
|
build-system = [
|
|
@@ -68,6 +68,7 @@ buildPythonPackage (finalAttrs: {
|
|
|
68
68
|
];
|
|
69
69
|
|
|
70
70
|
disabledTestPaths = [
|
|
71
|
+
# Requires geomstats which is not packaged
|
|
71
72
|
"tests/test_riemannian_neural_gas.py"
|
|
72
73
|
];
|
|
73
74
|
|
|
@@ -77,11 +78,11 @@ buildPythonPackage (finalAttrs: {
|
|
|
77
78
|
"prosemble.core"
|
|
78
79
|
];
|
|
79
80
|
|
|
80
|
-
meta =
|
|
81
|
+
meta = {
|
|
81
82
|
description = "JAX-based prototype machine learning library";
|
|
82
83
|
homepage = "https://github.com/naotoo1/prosemble";
|
|
83
84
|
changelog = "https://github.com/naotoo1/prosemble/releases/tag/v${finalAttrs.version}";
|
|
84
|
-
license = licenses.mit;
|
|
85
|
-
maintainers = with maintainers; [ ];
|
|
85
|
+
license = lib.licenses.mit;
|
|
86
|
+
maintainers = with lib.maintainers; [ naotoo1 ];
|
|
86
87
|
};
|
|
87
88
|
})
|
|
@@ -41,6 +41,19 @@ from .protocols import (
|
|
|
41
41
|
from .serialization import SerializationMixin
|
|
42
42
|
from .data import shuffle_arrays, padded_batches, batched_iterator
|
|
43
43
|
from .distributed import create_mesh, shard_data, replicate_params
|
|
44
|
+
from .optimizers import per_group_clip, hypergradient_descent, riemannian_nesterov
|
|
45
|
+
from .regularization import (
|
|
46
|
+
prototype_diversity_loss, prototype_diversity_loss_vectorized,
|
|
47
|
+
sparse_relevance_proximal, elastic_net_proximal,
|
|
48
|
+
)
|
|
49
|
+
from .curriculum import (
|
|
50
|
+
curriculum_weights, curriculum_threshold, apply_curriculum_to_loss,
|
|
51
|
+
)
|
|
52
|
+
from .reject import RejectOptionMixin
|
|
53
|
+
from .geodesic import (
|
|
54
|
+
geodesic_interpolation, geodesic_midpoint, decision_boundary_point,
|
|
55
|
+
prototype_geodesic_distances, inter_class_geodesics,
|
|
56
|
+
)
|
|
44
57
|
|
|
45
58
|
try:
|
|
46
59
|
import jax
|
|
@@ -93,6 +106,18 @@ __all__ = _distance_all + [
|
|
|
93
106
|
'shuffle_arrays', 'padded_batches', 'batched_iterator',
|
|
94
107
|
# Distributed training
|
|
95
108
|
'create_mesh', 'shard_data', 'replicate_params',
|
|
109
|
+
# Custom optimizers
|
|
110
|
+
'per_group_clip', 'hypergradient_descent', 'riemannian_nesterov',
|
|
111
|
+
# Regularization
|
|
112
|
+
'prototype_diversity_loss', 'prototype_diversity_loss_vectorized',
|
|
113
|
+
'sparse_relevance_proximal', 'elastic_net_proximal',
|
|
114
|
+
# Curriculum learning
|
|
115
|
+
'curriculum_weights', 'curriculum_threshold', 'apply_curriculum_to_loss',
|
|
116
|
+
# Reject option
|
|
117
|
+
'RejectOptionMixin',
|
|
118
|
+
# Geodesic interpolation
|
|
119
|
+
'geodesic_interpolation', 'geodesic_midpoint', 'decision_boundary_point',
|
|
120
|
+
'prototype_geodesic_distances', 'inter_class_geodesics',
|
|
96
121
|
# ONNX export (optional)
|
|
97
122
|
'export_onnx',
|
|
98
123
|
]
|
|
@@ -0,0 +1,171 @@
|
|
|
1
|
+
"""Curriculum learning (self-paced learning) for prototype-based models.
|
|
2
|
+
|
|
3
|
+
Implements the self-paced learning framework (Kumar et al. 2010) adapted
|
|
4
|
+
for LVQ models. Samples are presented in order of difficulty — easy
|
|
5
|
+
samples first, hard samples later.
|
|
6
|
+
|
|
7
|
+
For LVQ, sample difficulty is measured by the absolute value of the
|
|
8
|
+
mu-ratio: |mu(x)| close to 0 means the sample is on the decision
|
|
9
|
+
boundary (hard), while |mu(x)| >> 0 means it's far from the boundary
|
|
10
|
+
(easy).
|
|
11
|
+
|
|
12
|
+
Usage
|
|
13
|
+
-----
|
|
14
|
+
Curriculum learning is applied via sample masking in the loss function.
|
|
15
|
+
At each training step, a difficulty threshold lambda_t determines which
|
|
16
|
+
samples participate. The threshold increases over training to gradually
|
|
17
|
+
include harder samples.
|
|
18
|
+
|
|
19
|
+
References
|
|
20
|
+
----------
|
|
21
|
+
.. [1] Kumar, M. P., Packer, B., & Koller, D. (2010). Self-paced
|
|
22
|
+
learning for latent variable models. NeurIPS.
|
|
23
|
+
.. [2] Bengio, Y., Louradour, J., Collobert, R., & Weston, J. (2009).
|
|
24
|
+
Curriculum learning. ICML.
|
|
25
|
+
"""
|
|
26
|
+
|
|
27
|
+
import jax.numpy as jnp
|
|
28
|
+
|
|
29
|
+
|
|
30
|
+
def curriculum_weights(per_sample_loss, threshold, mode='hard'):
|
|
31
|
+
"""Compute per-sample curriculum weights based on loss magnitude.
|
|
32
|
+
|
|
33
|
+
Parameters
|
|
34
|
+
----------
|
|
35
|
+
per_sample_loss : array of shape (n_samples,)
|
|
36
|
+
Per-sample loss values (e.g., from GLVQ mu-ratio).
|
|
37
|
+
threshold : float
|
|
38
|
+
Difficulty threshold lambda_t. Samples with loss below this
|
|
39
|
+
threshold are included. Increases over training.
|
|
40
|
+
mode : str, {'hard', 'soft', 'linear'}
|
|
41
|
+
Weight assignment mode:
|
|
42
|
+
- 'hard': binary weights (0 or 1). Sample is either in or out.
|
|
43
|
+
- 'soft': smooth exponential weighting.
|
|
44
|
+
w_i = exp(-loss_i / threshold) if loss_i > threshold, else 1.
|
|
45
|
+
- 'linear': linear ramp-down.
|
|
46
|
+
w_i = max(0, 1 - (loss_i - threshold) / threshold).
|
|
47
|
+
Default: 'hard'.
|
|
48
|
+
|
|
49
|
+
Returns
|
|
50
|
+
-------
|
|
51
|
+
weights : array of shape (n_samples,)
|
|
52
|
+
Per-sample weights in [0, 1]. Weights sum is used to normalize
|
|
53
|
+
the loss.
|
|
54
|
+
"""
|
|
55
|
+
if mode == 'hard':
|
|
56
|
+
return jnp.where(per_sample_loss <= threshold, 1.0, 0.0)
|
|
57
|
+
elif mode == 'soft':
|
|
58
|
+
# Smooth transition: full weight below threshold, exponential decay above
|
|
59
|
+
excess = jnp.maximum(per_sample_loss - threshold, 0.0)
|
|
60
|
+
return jnp.exp(-excess / (threshold + 1e-10))
|
|
61
|
+
elif mode == 'linear':
|
|
62
|
+
# Linear ramp-down above threshold
|
|
63
|
+
return jnp.clip(1.0 - (per_sample_loss - threshold) / (threshold + 1e-10), 0.0, 1.0)
|
|
64
|
+
else:
|
|
65
|
+
raise ValueError(f"Unknown curriculum mode '{mode}'. Use 'hard', 'soft', or 'linear'.")
|
|
66
|
+
|
|
67
|
+
|
|
68
|
+
def curriculum_threshold(iteration, max_iter, init_threshold=0.3,
|
|
69
|
+
final_threshold=None, schedule='linear'):
|
|
70
|
+
"""Compute the difficulty threshold at a given training iteration.
|
|
71
|
+
|
|
72
|
+
The threshold increases over training: initially only easy samples
|
|
73
|
+
are included, and harder samples are gradually added.
|
|
74
|
+
|
|
75
|
+
Parameters
|
|
76
|
+
----------
|
|
77
|
+
iteration : int or array
|
|
78
|
+
Current training iteration.
|
|
79
|
+
max_iter : int
|
|
80
|
+
Maximum number of iterations.
|
|
81
|
+
init_threshold : float
|
|
82
|
+
Initial threshold (fraction of max loss to include). At the
|
|
83
|
+
start, only samples with loss < init_threshold * max_loss are
|
|
84
|
+
included. Default: 0.3.
|
|
85
|
+
final_threshold : float, optional
|
|
86
|
+
Final threshold. Default: None (uses a large value to include
|
|
87
|
+
all samples by end of training).
|
|
88
|
+
schedule : str, {'linear', 'exponential', 'cosine'}
|
|
89
|
+
How the threshold grows over training:
|
|
90
|
+
- 'linear': lambda_t = init + (final - init) * t / T
|
|
91
|
+
- 'exponential': lambda_t = init * (final / init) ^ (t / T)
|
|
92
|
+
- 'cosine': cosine annealing from init to final.
|
|
93
|
+
Default: 'linear'.
|
|
94
|
+
|
|
95
|
+
Returns
|
|
96
|
+
-------
|
|
97
|
+
threshold : float
|
|
98
|
+
Difficulty threshold at the current iteration.
|
|
99
|
+
"""
|
|
100
|
+
if final_threshold is None:
|
|
101
|
+
final_threshold = 100.0 # Effectively include all samples
|
|
102
|
+
|
|
103
|
+
progress = jnp.clip(iteration / jnp.maximum(max_iter - 1, 1), 0.0, 1.0)
|
|
104
|
+
|
|
105
|
+
if schedule == 'linear':
|
|
106
|
+
return init_threshold + (final_threshold - init_threshold) * progress
|
|
107
|
+
elif schedule == 'exponential':
|
|
108
|
+
# Avoid log(0): ensure init > 0
|
|
109
|
+
log_ratio = jnp.log(final_threshold / (init_threshold + 1e-10))
|
|
110
|
+
return init_threshold * jnp.exp(log_ratio * progress)
|
|
111
|
+
elif schedule == 'cosine':
|
|
112
|
+
# Cosine annealing: starts slow, accelerates, then slows
|
|
113
|
+
cosine_factor = 0.5 * (1.0 - jnp.cos(jnp.pi * progress))
|
|
114
|
+
return init_threshold + (final_threshold - init_threshold) * cosine_factor
|
|
115
|
+
else:
|
|
116
|
+
raise ValueError(
|
|
117
|
+
f"Unknown schedule '{schedule}'. Use 'linear', 'exponential', or 'cosine'."
|
|
118
|
+
)
|
|
119
|
+
|
|
120
|
+
|
|
121
|
+
def apply_curriculum_to_loss(per_sample_losses, iteration, max_iter,
|
|
122
|
+
init_percentile=0.3, schedule='linear',
|
|
123
|
+
mode='soft'):
|
|
124
|
+
"""Full curriculum pipeline: compute threshold and weighted loss.
|
|
125
|
+
|
|
126
|
+
Convenience function that combines threshold scheduling and weight
|
|
127
|
+
computation. Use inside ``_compute_loss`` to add curriculum learning
|
|
128
|
+
to any model.
|
|
129
|
+
|
|
130
|
+
Parameters
|
|
131
|
+
----------
|
|
132
|
+
per_sample_losses : array of shape (n_samples,)
|
|
133
|
+
Per-sample loss values.
|
|
134
|
+
iteration : int or array
|
|
135
|
+
Current training step.
|
|
136
|
+
max_iter : int
|
|
137
|
+
Total training steps.
|
|
138
|
+
init_percentile : float
|
|
139
|
+
Initial fraction of samples to include (by loss quantile).
|
|
140
|
+
Default: 0.3 (include the easiest 30% of samples initially).
|
|
141
|
+
schedule : str
|
|
142
|
+
Threshold growth schedule. Default: 'linear'.
|
|
143
|
+
mode : str
|
|
144
|
+
Weighting mode. Default: 'soft'.
|
|
145
|
+
|
|
146
|
+
Returns
|
|
147
|
+
-------
|
|
148
|
+
weighted_loss : scalar
|
|
149
|
+
Curriculum-weighted mean loss.
|
|
150
|
+
"""
|
|
151
|
+
# Estimate threshold from loss statistics
|
|
152
|
+
# Use quantile of current losses as adaptive threshold
|
|
153
|
+
sorted_losses = jnp.sort(per_sample_losses)
|
|
154
|
+
n = per_sample_losses.shape[0]
|
|
155
|
+
|
|
156
|
+
# Progress determines what fraction of samples to include
|
|
157
|
+
progress = jnp.clip(iteration / jnp.maximum(max_iter - 1, 1), 0.0, 1.0)
|
|
158
|
+
# Start with init_percentile fraction, end with all samples
|
|
159
|
+
include_frac = init_percentile + (1.0 - init_percentile) * progress
|
|
160
|
+
|
|
161
|
+
# Convert fraction to threshold via quantile
|
|
162
|
+
idx = jnp.clip((include_frac * n).astype(jnp.int32), 0, n - 1)
|
|
163
|
+
threshold = sorted_losses[idx]
|
|
164
|
+
|
|
165
|
+
weights = curriculum_weights(per_sample_losses, threshold, mode=mode)
|
|
166
|
+
|
|
167
|
+
# Weighted mean loss (avoid division by zero)
|
|
168
|
+
total_weight = jnp.sum(weights)
|
|
169
|
+
weighted_loss = jnp.sum(weights * per_sample_losses) / (total_weight + 1e-10)
|
|
170
|
+
|
|
171
|
+
return weighted_loss
|
|
@@ -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',
|
|
@@ -0,0 +1,217 @@
|
|
|
1
|
+
"""Geodesic interpolation and boundary visualization for Riemannian models.
|
|
2
|
+
|
|
3
|
+
Computes geodesic paths between prototypes on Riemannian manifolds for
|
|
4
|
+
decision boundary visualization and model interpretation.
|
|
5
|
+
|
|
6
|
+
On curved manifolds, the decision boundary (locus where d(x, w_i) = d(x, w_j))
|
|
7
|
+
is not a hyperplane — it's a curved surface. Geodesic interpolation allows
|
|
8
|
+
us to:
|
|
9
|
+
1. Visualize the path between prototypes along the manifold.
|
|
10
|
+
2. Find the approximate decision boundary point along geodesics.
|
|
11
|
+
3. Compute prototype midpoints respecting manifold geometry.
|
|
12
|
+
|
|
13
|
+
References
|
|
14
|
+
----------
|
|
15
|
+
.. [1] Absil, P.-A., Mahony, R., & Sepulchre, R. (2008). Optimization
|
|
16
|
+
Algorithms on Matrix Manifolds. Princeton University Press.
|
|
17
|
+
"""
|
|
18
|
+
|
|
19
|
+
import jax.numpy as jnp
|
|
20
|
+
|
|
21
|
+
|
|
22
|
+
def geodesic_interpolation(manifold, point_a, point_b, n_points=50):
|
|
23
|
+
"""Compute a geodesic path between two points on a manifold.
|
|
24
|
+
|
|
25
|
+
Uses the exponential map to trace the shortest path (geodesic)
|
|
26
|
+
between two manifold points:
|
|
27
|
+
|
|
28
|
+
.. math::
|
|
29
|
+
|
|
30
|
+
\\gamma(t) = \\text{Exp}_{w_a}(t \\cdot \\text{Log}_{w_a}(w_b)),
|
|
31
|
+
\\quad t \\in [0, 1]
|
|
32
|
+
|
|
33
|
+
Parameters
|
|
34
|
+
----------
|
|
35
|
+
manifold : SO, SPD, or Grassmannian
|
|
36
|
+
Riemannian manifold instance with ``exp_map`` and ``log_map``.
|
|
37
|
+
point_a : array
|
|
38
|
+
Starting point on the manifold.
|
|
39
|
+
point_b : array
|
|
40
|
+
End point on the manifold.
|
|
41
|
+
n_points : int
|
|
42
|
+
Number of interpolation points along the geodesic.
|
|
43
|
+
Default: 50.
|
|
44
|
+
|
|
45
|
+
Returns
|
|
46
|
+
-------
|
|
47
|
+
path : array of shape (n_points, ...)
|
|
48
|
+
Points along the geodesic from point_a to point_b.
|
|
49
|
+
Each point lies on the manifold.
|
|
50
|
+
"""
|
|
51
|
+
# Compute the tangent vector from a to b
|
|
52
|
+
tangent = manifold.log_map(point_a, point_b)
|
|
53
|
+
|
|
54
|
+
# Interpolate along the geodesic
|
|
55
|
+
t_values = jnp.linspace(0.0, 1.0, n_points)
|
|
56
|
+
path = []
|
|
57
|
+
for t in t_values:
|
|
58
|
+
point = manifold.exp_map(point_a, t * tangent)
|
|
59
|
+
path.append(point)
|
|
60
|
+
|
|
61
|
+
return jnp.stack(path, axis=0)
|
|
62
|
+
|
|
63
|
+
|
|
64
|
+
def geodesic_midpoint(manifold, point_a, point_b):
|
|
65
|
+
"""Compute the geodesic midpoint between two manifold points.
|
|
66
|
+
|
|
67
|
+
The midpoint is at t = 0.5 along the geodesic:
|
|
68
|
+
|
|
69
|
+
.. math::
|
|
70
|
+
|
|
71
|
+
m = \\text{Exp}_{w_a}(0.5 \\cdot \\text{Log}_{w_a}(w_b))
|
|
72
|
+
|
|
73
|
+
Parameters
|
|
74
|
+
----------
|
|
75
|
+
manifold : SO, SPD, or Grassmannian
|
|
76
|
+
Riemannian manifold instance.
|
|
77
|
+
point_a : array
|
|
78
|
+
First manifold point.
|
|
79
|
+
point_b : array
|
|
80
|
+
Second manifold point.
|
|
81
|
+
|
|
82
|
+
Returns
|
|
83
|
+
-------
|
|
84
|
+
midpoint : array
|
|
85
|
+
Geodesic midpoint, guaranteed to lie on the manifold.
|
|
86
|
+
"""
|
|
87
|
+
tangent = manifold.log_map(point_a, point_b)
|
|
88
|
+
return manifold.exp_map(point_a, 0.5 * tangent)
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
def decision_boundary_point(manifold, proto_a, proto_b, n_search=100):
|
|
92
|
+
"""Find the decision boundary point along the geodesic between prototypes.
|
|
93
|
+
|
|
94
|
+
On curved manifolds, the equidistant point (where d(x, w_a) = d(x, w_b))
|
|
95
|
+
is not necessarily at t = 0.5. This function searches along the geodesic
|
|
96
|
+
for the point where distances to both prototypes are equal.
|
|
97
|
+
|
|
98
|
+
Parameters
|
|
99
|
+
----------
|
|
100
|
+
manifold : SO, SPD, or Grassmannian
|
|
101
|
+
Riemannian manifold instance.
|
|
102
|
+
proto_a : array
|
|
103
|
+
First prototype (on manifold).
|
|
104
|
+
proto_b : array
|
|
105
|
+
Second prototype (on manifold).
|
|
106
|
+
n_search : int
|
|
107
|
+
Number of candidate points to evaluate. Default: 100.
|
|
108
|
+
|
|
109
|
+
Returns
|
|
110
|
+
-------
|
|
111
|
+
boundary_point : array
|
|
112
|
+
Point on the geodesic where d(point, proto_a) = d(point, proto_b).
|
|
113
|
+
t_boundary : float
|
|
114
|
+
Parameter value t in [0, 1] where the boundary lies.
|
|
115
|
+
t = 0.5 for symmetric manifolds, may differ on curved spaces.
|
|
116
|
+
"""
|
|
117
|
+
tangent = manifold.log_map(proto_a, proto_b)
|
|
118
|
+
t_values = jnp.linspace(0.0, 1.0, n_search)
|
|
119
|
+
|
|
120
|
+
# Compute distance difference at each point along the geodesic
|
|
121
|
+
def distance_diff(t):
|
|
122
|
+
point = manifold.exp_map(proto_a, t * tangent)
|
|
123
|
+
d_a = manifold.distance(point, proto_a)
|
|
124
|
+
d_b = manifold.distance(point, proto_b)
|
|
125
|
+
return jnp.abs(d_a - d_b)
|
|
126
|
+
|
|
127
|
+
diffs = jnp.array([distance_diff(t) for t in t_values])
|
|
128
|
+
best_idx = jnp.argmin(diffs)
|
|
129
|
+
t_boundary = t_values[best_idx]
|
|
130
|
+
boundary_point = manifold.exp_map(proto_a, t_boundary * tangent)
|
|
131
|
+
|
|
132
|
+
return boundary_point, float(t_boundary)
|
|
133
|
+
|
|
134
|
+
|
|
135
|
+
def prototype_geodesic_distances(manifold, prototypes, proto_labels):
|
|
136
|
+
"""Compute pairwise geodesic distances between prototypes.
|
|
137
|
+
|
|
138
|
+
Parameters
|
|
139
|
+
----------
|
|
140
|
+
manifold : SO, SPD, or Grassmannian
|
|
141
|
+
Riemannian manifold instance.
|
|
142
|
+
prototypes : array of shape (n_prototypes, ...)
|
|
143
|
+
Prototype points on the manifold (flattened).
|
|
144
|
+
proto_labels : array of shape (n_prototypes,)
|
|
145
|
+
Class labels for prototypes.
|
|
146
|
+
|
|
147
|
+
Returns
|
|
148
|
+
-------
|
|
149
|
+
distances : array of shape (n_prototypes, n_prototypes)
|
|
150
|
+
Pairwise geodesic distance matrix.
|
|
151
|
+
"""
|
|
152
|
+
n = prototypes.shape[0]
|
|
153
|
+
distances = jnp.zeros((n, n))
|
|
154
|
+
|
|
155
|
+
for i in range(n):
|
|
156
|
+
for j in range(i + 1, n):
|
|
157
|
+
d = manifold.distance(prototypes[i], prototypes[j])
|
|
158
|
+
distances = distances.at[i, j].set(d)
|
|
159
|
+
distances = distances.at[j, i].set(d)
|
|
160
|
+
|
|
161
|
+
return distances
|
|
162
|
+
|
|
163
|
+
|
|
164
|
+
def inter_class_geodesics(manifold, prototypes, proto_labels, n_points=50):
|
|
165
|
+
"""Compute geodesic paths between all inter-class prototype pairs.
|
|
166
|
+
|
|
167
|
+
Useful for visualizing decision boundaries: the boundary between
|
|
168
|
+
two classes lies somewhere along the geodesic connecting their
|
|
169
|
+
closest prototypes.
|
|
170
|
+
|
|
171
|
+
Parameters
|
|
172
|
+
----------
|
|
173
|
+
manifold : SO, SPD, or Grassmannian
|
|
174
|
+
Riemannian manifold instance.
|
|
175
|
+
prototypes : array of shape (n_prototypes, ...)
|
|
176
|
+
Prototype positions on manifold.
|
|
177
|
+
proto_labels : array of shape (n_prototypes,)
|
|
178
|
+
Class labels.
|
|
179
|
+
n_points : int
|
|
180
|
+
Points per geodesic path. Default: 50.
|
|
181
|
+
|
|
182
|
+
Returns
|
|
183
|
+
-------
|
|
184
|
+
geodesics : list of dict
|
|
185
|
+
Each dict contains:
|
|
186
|
+
- 'path': array of shape (n_points, ...) — geodesic path
|
|
187
|
+
- 'proto_a_idx': int — index of first prototype
|
|
188
|
+
- 'proto_b_idx': int — index of second prototype
|
|
189
|
+
- 'class_a': int — class of first prototype
|
|
190
|
+
- 'class_b': int — class of second prototype
|
|
191
|
+
- 'boundary_t': float — approximate boundary location
|
|
192
|
+
"""
|
|
193
|
+
import numpy as np
|
|
194
|
+
|
|
195
|
+
n = prototypes.shape[0]
|
|
196
|
+
labels_np = np.asarray(proto_labels)
|
|
197
|
+
geodesics = []
|
|
198
|
+
|
|
199
|
+
for i in range(n):
|
|
200
|
+
for j in range(i + 1, n):
|
|
201
|
+
if labels_np[i] != labels_np[j]:
|
|
202
|
+
path = geodesic_interpolation(
|
|
203
|
+
manifold, prototypes[i], prototypes[j], n_points
|
|
204
|
+
)
|
|
205
|
+
_, t_boundary = decision_boundary_point(
|
|
206
|
+
manifold, prototypes[i], prototypes[j]
|
|
207
|
+
)
|
|
208
|
+
geodesics.append({
|
|
209
|
+
'path': path,
|
|
210
|
+
'proto_a_idx': i,
|
|
211
|
+
'proto_b_idx': j,
|
|
212
|
+
'class_a': int(labels_np[i]),
|
|
213
|
+
'class_b': int(labels_np[j]),
|
|
214
|
+
'boundary_t': t_boundary,
|
|
215
|
+
})
|
|
216
|
+
|
|
217
|
+
return geodesics
|