prosemble 2.1.0__tar.gz → 2.1.1__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- {prosemble-2.1.0 → prosemble-2.1.1}/.github/workflows/release.yml +1 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/PKG-INFO +1 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/__init__.py +3 -0
- prosemble-2.1.1/prosemble/core/distributed.py +104 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/glvq.py +3 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lvq1.py +3 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lvq21.py +3 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/median_lvq.py +3 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/probabilistic_lvq.py +6 -2
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/prototype_base.py +89 -10
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/PKG-INFO +1 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/SOURCES.txt +2 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/pyproject.toml +1 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/api/core.rst +7 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/conf.py +1 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/advanced.rst +122 -0
- prosemble-2.1.1/tests/test_distributed.py +249 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/uv.lock +1 -1
- {prosemble-2.1.0 → prosemble-2.1.1}/.bumpversion.cfg +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.editorconfig +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.envrc +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.github/ISSUE_TEMPLATE/config.yml +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.github/ISSUE_TEMPLATE/feature_request.md +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.github/workflows/ci.yml +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.gitignore +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/.readthedocs.yaml +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/CONSTITUTION.md +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/LICENSE +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/MANIFEST.in +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/ONNX_COVERAGE.md +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/README.md +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/TASK.md +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/benchmarks/benchmark_distance_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/benchmarks/benchmark_fcm_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/benchmarks/benchmark_pcm_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/devenv.lock +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/devenv.nix +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/devenv.yaml +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/afcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/bgpc_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/cbc_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/celvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/fcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/fpcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/glvq_advanced_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/glvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/glvq_unequal_prototypes_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/gmlvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/grlvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/hcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/heskes_som_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/ipcm2_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/ipcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kafcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kfcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kfpcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kipcm2_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kipcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kmeans_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/knn_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kohonen_som_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kpcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/kpfcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/large_dataset_training.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/lcelvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/lmrslvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/lmrslvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/lvq1_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/lvqmln_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/mcelvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/mrslvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/mrslvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/neural_gas_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/npc_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_lmrslvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_lmrslvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_mrslvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_mrslvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_rslvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_rslvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/pcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/pfcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/plvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/rslvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/slng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/smng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/som_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/srng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/stng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/tcelvq_ng_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_fcm_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_glvq_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_neural_gas_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_som_iris.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/flake.nix +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/mkdocs.yml +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/nix/package.nix +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/__init__.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/activations.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/callbacks.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/competitions.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/data.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/distance.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/initializers.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/kernel.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/losses.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/manifolds.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/model_selection.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/onnx_export.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/pipeline.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/pooling.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/protocols.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/quantization.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/serialization.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/similarities.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/utils.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/vis.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/visualizer.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/datasets/__init__.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/datasets/dataset.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/__init__.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/afcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/base.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/bgpc.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/cbc.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/celvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/celvq_ng_mixin.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/crossentropy_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/fcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/fpcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/growing_neural_gas.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/hcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/heskes_som.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/image_cbc.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/image_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/ipcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/ipcm2.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kafcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kfcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kfpcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kipcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kipcm2.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kmeans.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/knn.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kohonen_som.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kpcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kpfcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lcelvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/local_matrix_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lvqmln.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/matrix_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/matrix_rslvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/mcelvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/mrslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/neural_gas.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/npc.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_glvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_glvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_glvq_ng_mixin.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gmlvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gmlvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_grlvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_grlvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gtlvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gtlvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_lgmlvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_lgmlvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_mrslvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_mrslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_rslvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_rslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_rslvq_ng_mixin.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/pcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/pfcm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/plvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/relevance_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_neural_gas.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_slng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_smng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_srng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_stng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/rslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/siamese_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/slng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/smng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/som.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/srng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/stng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_lm.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_m.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_r.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_t.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/tangent_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/tcelvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/py.typed +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/dependency_links.txt +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/requires.txt +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/top_level.txt +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/requirements_docs.txt +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/setup.cfg +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/Makefile +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/api/models.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/clustering.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/one_class.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/onnx.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/riemannian.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/supervised.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/topology.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/index.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/installation.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/make.bat +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/quickstart.rst +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/__init__.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_activations.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_cbc.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_celvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_checkpoint.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_clustering_patience.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_competitions.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_core.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_core_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_data.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_dataset.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_dataset_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_distance_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_examples.sh +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_fcm_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_glvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_gmlvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_heskes_som.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_image_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_initializers.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_jax_models.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_jit_inference.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_kohonen_som.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_lcelvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_losses.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_lvq_nongradient.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_lvqmln_plvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_matrix_rslvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_mcelvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_minibatch.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_mixed_precision.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_model_selection.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_mrslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_neural_gas.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_glvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_glvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_glvq_variants.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_mrslvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_rslvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_rslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_onnx_export.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pcm_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pfcm_jax.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pipeline.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pooling.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_probabilistic.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_protocols.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_prototype_base.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_quantization.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_riemannian_neural_gas.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_riemannian_supervised_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_rslvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_scan_training.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_serialization.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_siamese_lvq.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_similarities.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_slng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_smng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_srng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_stng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_svq_occ.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_svq_occ_r.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_svq_occ_variants.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_tcelvq_ng.py +0 -0
- {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_vis.py +0 -0
|
@@ -24,7 +24,7 @@ jobs:
|
|
|
24
24
|
python -m pip install --upgrade pip
|
|
25
25
|
pip install .[all]
|
|
26
26
|
- name: Run tests
|
|
27
|
-
run: pytest tests/ -x -q --ignore=tests/test_oc_glvq.py --ignore=tests/test_oc_glvq_ng.py --ignore=tests/test_oc_glvq_variants.py --ignore=tests/test_svq_occ.py --ignore=tests/test_svq_occ_r.py --ignore=tests/test_svq_occ_variants.py --ignore=tests/test_smng.py --ignore=tests/test_slng.py --ignore=tests/test_stng.py --ignore=tests/test_riemannian_neural_gas.py
|
|
27
|
+
run: pytest tests/ -x -q --ignore=tests/test_oc_glvq.py --ignore=tests/test_oc_glvq_ng.py --ignore=tests/test_oc_glvq_variants.py --ignore=tests/test_svq_occ.py --ignore=tests/test_svq_occ_r.py --ignore=tests/test_svq_occ_variants.py --ignore=tests/test_smng.py --ignore=tests/test_slng.py --ignore=tests/test_stng.py --ignore=tests/test_riemannian_neural_gas.py --ignore=tests/test_riemannian_supervised_ng.py
|
|
28
28
|
|
|
29
29
|
build:
|
|
30
30
|
needs: test
|
|
@@ -40,6 +40,7 @@ from .protocols import (
|
|
|
40
40
|
)
|
|
41
41
|
from .serialization import SerializationMixin
|
|
42
42
|
from .data import shuffle_arrays, padded_batches, batched_iterator
|
|
43
|
+
from .distributed import create_mesh, shard_data, replicate_params
|
|
43
44
|
|
|
44
45
|
try:
|
|
45
46
|
import jax
|
|
@@ -90,6 +91,8 @@ __all__ = _distance_all + [
|
|
|
90
91
|
'SerializationMixin',
|
|
91
92
|
# Data loading utilities
|
|
92
93
|
'shuffle_arrays', 'padded_batches', 'batched_iterator',
|
|
94
|
+
# Distributed training
|
|
95
|
+
'create_mesh', 'shard_data', 'replicate_params',
|
|
93
96
|
# ONNX export (optional)
|
|
94
97
|
'export_onnx',
|
|
95
98
|
]
|
|
@@ -0,0 +1,104 @@
|
|
|
1
|
+
"""Distributed training utilities for multi-device data parallelism.
|
|
2
|
+
|
|
3
|
+
Provides functions for sharding data across devices and replicating
|
|
4
|
+
model parameters, enabling data-parallel training on multi-GPU/TPU setups.
|
|
5
|
+
"""
|
|
6
|
+
|
|
7
|
+
import jax
|
|
8
|
+
import jax.numpy as jnp
|
|
9
|
+
from jax.sharding import Mesh, NamedSharding, PartitionSpec as P
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
def create_mesh(devices=None):
|
|
13
|
+
"""Create a 1D device mesh for data parallelism.
|
|
14
|
+
|
|
15
|
+
Parameters
|
|
16
|
+
----------
|
|
17
|
+
devices : list of jax.Device or None
|
|
18
|
+
Devices to use. If None, returns None (single-device mode).
|
|
19
|
+
|
|
20
|
+
Returns
|
|
21
|
+
-------
|
|
22
|
+
Mesh or None
|
|
23
|
+
"""
|
|
24
|
+
if devices is None:
|
|
25
|
+
return None
|
|
26
|
+
return Mesh(devices, axis_names=('data',))
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
def shard_data(X, y, mesh):
|
|
30
|
+
"""Shard data arrays along the batch dimension across devices.
|
|
31
|
+
|
|
32
|
+
Parameters
|
|
33
|
+
----------
|
|
34
|
+
X : jnp.ndarray of shape (n_samples, ...)
|
|
35
|
+
Input data.
|
|
36
|
+
y : jnp.ndarray of shape (n_samples,)
|
|
37
|
+
Labels.
|
|
38
|
+
mesh : Mesh
|
|
39
|
+
Device mesh from create_mesh().
|
|
40
|
+
|
|
41
|
+
Returns
|
|
42
|
+
-------
|
|
43
|
+
X_sharded, y_sharded : tuple of sharded arrays
|
|
44
|
+
"""
|
|
45
|
+
data_sharding = NamedSharding(mesh, P('data'))
|
|
46
|
+
X_sharded = jax.device_put(X, data_sharding)
|
|
47
|
+
y_sharded = jax.device_put(y, data_sharding)
|
|
48
|
+
return X_sharded, y_sharded
|
|
49
|
+
|
|
50
|
+
|
|
51
|
+
def replicate_params(params, mesh):
|
|
52
|
+
"""Replicate params across all devices (no partitioning).
|
|
53
|
+
|
|
54
|
+
Parameters
|
|
55
|
+
----------
|
|
56
|
+
params : dict (pytree)
|
|
57
|
+
Model parameters.
|
|
58
|
+
mesh : Mesh
|
|
59
|
+
Device mesh from create_mesh().
|
|
60
|
+
|
|
61
|
+
Returns
|
|
62
|
+
-------
|
|
63
|
+
params replicated across mesh
|
|
64
|
+
"""
|
|
65
|
+
replicated = NamedSharding(mesh, P())
|
|
66
|
+
return jax.device_put(params, replicated)
|
|
67
|
+
|
|
68
|
+
|
|
69
|
+
def replicate_opt_state(opt_state, mesh):
|
|
70
|
+
"""Replicate optimizer state across all devices.
|
|
71
|
+
|
|
72
|
+
Parameters
|
|
73
|
+
----------
|
|
74
|
+
opt_state : pytree
|
|
75
|
+
Optax optimizer state.
|
|
76
|
+
mesh : Mesh
|
|
77
|
+
Device mesh from create_mesh().
|
|
78
|
+
|
|
79
|
+
Returns
|
|
80
|
+
-------
|
|
81
|
+
opt_state replicated across mesh
|
|
82
|
+
"""
|
|
83
|
+
replicated = NamedSharding(mesh, P())
|
|
84
|
+
return jax.device_put(opt_state, replicated)
|
|
85
|
+
|
|
86
|
+
|
|
87
|
+
def unshard_params(params):
|
|
88
|
+
"""Bring params back to a single device.
|
|
89
|
+
|
|
90
|
+
Used after training to store results as plain arrays for
|
|
91
|
+
predict/export operations.
|
|
92
|
+
|
|
93
|
+
Parameters
|
|
94
|
+
----------
|
|
95
|
+
params : pytree
|
|
96
|
+
Potentially sharded parameters.
|
|
97
|
+
|
|
98
|
+
Returns
|
|
99
|
+
-------
|
|
100
|
+
params on default device
|
|
101
|
+
"""
|
|
102
|
+
return jax.tree.map(
|
|
103
|
+
lambda x: jax.device_put(x, jax.devices()[0]), params
|
|
104
|
+
)
|
|
@@ -117,7 +117,8 @@ class GLVQ(SupervisedPrototypeModel):
|
|
|
117
117
|
prototypes_initializer=None, patience=None,
|
|
118
118
|
restore_best=False, class_weight=None,
|
|
119
119
|
gradient_accumulation_steps=None, ema_decay=None,
|
|
120
|
-
freeze_params=None, lookahead=None, mixed_precision=None
|
|
120
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
121
|
+
**kwargs):
|
|
121
122
|
super().__init__(
|
|
122
123
|
n_prototypes_per_class=n_prototypes_per_class,
|
|
123
124
|
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
@@ -132,6 +133,7 @@ class GLVQ(SupervisedPrototypeModel):
|
|
|
132
133
|
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
133
134
|
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
134
135
|
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
136
|
+
**kwargs,
|
|
135
137
|
)
|
|
136
138
|
self.beta = beta
|
|
137
139
|
|
|
@@ -117,7 +117,8 @@ class LVQ1(SupervisedPrototypeModel):
|
|
|
117
117
|
prototypes_initializer=None, patience=None,
|
|
118
118
|
restore_best=False, class_weight=None,
|
|
119
119
|
gradient_accumulation_steps=None, ema_decay=None,
|
|
120
|
-
freeze_params=None, lookahead=None, mixed_precision=None
|
|
120
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
121
|
+
**kwargs):
|
|
121
122
|
super().__init__(
|
|
122
123
|
n_prototypes_per_class=n_prototypes_per_class,
|
|
123
124
|
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
@@ -132,6 +133,7 @@ class LVQ1(SupervisedPrototypeModel):
|
|
|
132
133
|
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
133
134
|
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
134
135
|
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
136
|
+
**kwargs,
|
|
135
137
|
)
|
|
136
138
|
|
|
137
139
|
def fit(self, X, y, initial_prototypes=None):
|
|
@@ -116,7 +116,8 @@ class LVQ21(SupervisedPrototypeModel):
|
|
|
116
116
|
prototypes_initializer=None, patience=None,
|
|
117
117
|
restore_best=False, class_weight=None,
|
|
118
118
|
gradient_accumulation_steps=None, ema_decay=None,
|
|
119
|
-
freeze_params=None, lookahead=None, mixed_precision=None
|
|
119
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
120
|
+
**kwargs):
|
|
120
121
|
super().__init__(
|
|
121
122
|
n_prototypes_per_class=n_prototypes_per_class,
|
|
122
123
|
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
@@ -131,6 +132,7 @@ class LVQ21(SupervisedPrototypeModel):
|
|
|
131
132
|
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
132
133
|
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
133
134
|
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
135
|
+
**kwargs,
|
|
134
136
|
)
|
|
135
137
|
|
|
136
138
|
def fit(self, X, y, initial_prototypes=None):
|
|
@@ -117,7 +117,8 @@ class MedianLVQ(SupervisedPrototypeModel):
|
|
|
117
117
|
prototypes_initializer=None, patience=None,
|
|
118
118
|
restore_best=False, class_weight=None,
|
|
119
119
|
gradient_accumulation_steps=None, ema_decay=None,
|
|
120
|
-
freeze_params=None, lookahead=None, mixed_precision=None
|
|
120
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
121
|
+
**kwargs):
|
|
121
122
|
super().__init__(
|
|
122
123
|
n_prototypes_per_class=n_prototypes_per_class,
|
|
123
124
|
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
@@ -132,6 +133,7 @@ class MedianLVQ(SupervisedPrototypeModel):
|
|
|
132
133
|
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
133
134
|
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
134
135
|
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
136
|
+
**kwargs,
|
|
135
137
|
)
|
|
136
138
|
|
|
137
139
|
def fit(self, X, y, initial_prototypes=None):
|
|
@@ -130,7 +130,8 @@ class SLVQ(SupervisedPrototypeModel):
|
|
|
130
130
|
prototypes_initializer=None, patience=None,
|
|
131
131
|
restore_best=False, class_weight=None,
|
|
132
132
|
gradient_accumulation_steps=None, ema_decay=None,
|
|
133
|
-
freeze_params=None, lookahead=None, mixed_precision=None
|
|
133
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
134
|
+
**kwargs):
|
|
134
135
|
super().__init__(
|
|
135
136
|
n_prototypes_per_class=n_prototypes_per_class,
|
|
136
137
|
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
@@ -145,6 +146,7 @@ class SLVQ(SupervisedPrototypeModel):
|
|
|
145
146
|
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
146
147
|
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
147
148
|
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
149
|
+
**kwargs,
|
|
148
150
|
)
|
|
149
151
|
self.sigma = sigma
|
|
150
152
|
self.rejection_confidence = rejection_confidence
|
|
@@ -291,7 +293,8 @@ class RSLVQ(SupervisedPrototypeModel):
|
|
|
291
293
|
prototypes_initializer=None, patience=None,
|
|
292
294
|
restore_best=False, class_weight=None,
|
|
293
295
|
gradient_accumulation_steps=None, ema_decay=None,
|
|
294
|
-
freeze_params=None, lookahead=None, mixed_precision=None
|
|
296
|
+
freeze_params=None, lookahead=None, mixed_precision=None,
|
|
297
|
+
**kwargs):
|
|
295
298
|
super().__init__(
|
|
296
299
|
n_prototypes_per_class=n_prototypes_per_class,
|
|
297
300
|
max_iter=max_iter, lr=lr, epsilon=epsilon,
|
|
@@ -306,6 +309,7 @@ class RSLVQ(SupervisedPrototypeModel):
|
|
|
306
309
|
gradient_accumulation_steps=gradient_accumulation_steps,
|
|
307
310
|
ema_decay=ema_decay, freeze_params=freeze_params,
|
|
308
311
|
lookahead=lookahead, mixed_precision=mixed_precision,
|
|
312
|
+
**kwargs,
|
|
309
313
|
)
|
|
310
314
|
self.sigma = sigma
|
|
311
315
|
self.rejection_confidence = rejection_confidence
|
|
@@ -246,6 +246,16 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
246
246
|
Master weights stay in float32; forward/backward pass runs in lower
|
|
247
247
|
precision for ~2x speed and ~half memory on GPU. Float16 uses static
|
|
248
248
|
loss scaling to prevent gradient underflow. Default: None (disabled).
|
|
249
|
+
gradient_checkpointing : bool, optional
|
|
250
|
+
If True, applies ``jax.checkpoint`` (remat) to the loss function.
|
|
251
|
+
Trades compute for memory by re-computing forward activations during
|
|
252
|
+
the backward pass. Beneficial for deep backbone models (Image,
|
|
253
|
+
Siamese). Default: False.
|
|
254
|
+
devices : list of jax.Device or None, optional
|
|
255
|
+
Devices for data-parallel training. Data is sharded across devices
|
|
256
|
+
along the batch dimension; params and optimizer state are replicated.
|
|
257
|
+
When using mini-batch training, ``batch_size`` must be divisible by
|
|
258
|
+
the number of devices. Default: None (single-device).
|
|
249
259
|
"""
|
|
250
260
|
|
|
251
261
|
def __init__(
|
|
@@ -273,6 +283,8 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
273
283
|
freeze_params: list | None = None,
|
|
274
284
|
lookahead: dict | None = None,
|
|
275
285
|
mixed_precision: str | None = None,
|
|
286
|
+
gradient_checkpointing: bool = False,
|
|
287
|
+
devices: list | None = None,
|
|
276
288
|
):
|
|
277
289
|
if isinstance(n_prototypes_per_class, dict):
|
|
278
290
|
for cls, cnt in n_prototypes_per_class.items():
|
|
@@ -356,6 +368,13 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
356
368
|
self._mp_dtype = jnp.dtype(mixed_precision) if mixed_precision else None
|
|
357
369
|
self._loss_scale = 2**15 if mixed_precision == 'float16' else 1.0
|
|
358
370
|
|
|
371
|
+
# Gradient checkpointing (remat)
|
|
372
|
+
self.gradient_checkpointing = gradient_checkpointing
|
|
373
|
+
|
|
374
|
+
# Multi-device data parallelism
|
|
375
|
+
self.devices = devices
|
|
376
|
+
self._mesh = None
|
|
377
|
+
|
|
359
378
|
# Default active optimizer (may be wrapped with masking in fit())
|
|
360
379
|
self._active_optimizer = self._optimizer
|
|
361
380
|
|
|
@@ -653,17 +672,31 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
653
672
|
return self._compute_loss(params_mp, X, y, proto_labels) * self._loss_scale
|
|
654
673
|
|
|
655
674
|
def _value_and_grad_fn(self, params, X, y, proto_labels):
|
|
656
|
-
"""Compute loss and gradients
|
|
657
|
-
|
|
658
|
-
|
|
659
|
-
|
|
660
|
-
|
|
661
|
-
|
|
675
|
+
"""Compute loss and gradients.
|
|
676
|
+
|
|
677
|
+
Supports optional mixed precision, gradient checkpointing (remat),
|
|
678
|
+
and custom VJP rules.
|
|
679
|
+
"""
|
|
680
|
+
# Select loss function
|
|
681
|
+
if getattr(self, '_use_custom_vjp', False):
|
|
682
|
+
loss_fn = self._custom_vjp_loss
|
|
683
|
+
elif self._mp_dtype is not None:
|
|
684
|
+
loss_fn = self._compute_loss_mp
|
|
685
|
+
else:
|
|
686
|
+
loss_fn = self._compute_loss
|
|
687
|
+
|
|
688
|
+
# Apply gradient checkpointing (recompute forward during backward)
|
|
689
|
+
if self.gradient_checkpointing and not getattr(self, '_use_custom_vjp', False):
|
|
690
|
+
loss_fn = jax.checkpoint(loss_fn)
|
|
691
|
+
|
|
692
|
+
loss, grads = jax.value_and_grad(loss_fn)(params, X, y, proto_labels)
|
|
693
|
+
|
|
694
|
+
# Unscale for mixed precision
|
|
695
|
+
if self._mp_dtype is not None and not getattr(self, '_use_custom_vjp', False):
|
|
696
|
+
loss = loss / self._loss_scale
|
|
662
697
|
grads = jax.tree.map(lambda g: g / self._loss_scale, grads)
|
|
663
|
-
|
|
664
|
-
return
|
|
665
|
-
params, X, y, proto_labels
|
|
666
|
-
)
|
|
698
|
+
|
|
699
|
+
return loss, grads
|
|
667
700
|
|
|
668
701
|
def _post_update(self, params):
|
|
669
702
|
"""Apply constraints after gradient update (e.g., normalize relevances).
|
|
@@ -672,6 +705,16 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
672
705
|
"""
|
|
673
706
|
return params
|
|
674
707
|
|
|
708
|
+
def _custom_vjp_loss(self, params, X, y, proto_labels):
|
|
709
|
+
"""Compute loss with custom VJP rule.
|
|
710
|
+
|
|
711
|
+
Override this method and set ``self._use_custom_vjp = True`` in
|
|
712
|
+
``__init__`` to use a custom backward pass via ``@jax.custom_vjp``.
|
|
713
|
+
|
|
714
|
+
Default: None (uses standard autodiff on ``_compute_loss``).
|
|
715
|
+
"""
|
|
716
|
+
return None
|
|
717
|
+
|
|
675
718
|
def _get_weighted_data(self, X, y, sample_weight, key):
|
|
676
719
|
"""Apply sample weighting via weighted resampling.
|
|
677
720
|
|
|
@@ -749,6 +792,13 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
749
792
|
|
|
750
793
|
if self.restore_best and best_params is not None:
|
|
751
794
|
params = best_params
|
|
795
|
+
|
|
796
|
+
# Unshard params back to single device after multi-device training
|
|
797
|
+
if self._mesh is not None:
|
|
798
|
+
from prosemble.core.distributed import unshard_params
|
|
799
|
+
params = unshard_params(params)
|
|
800
|
+
proto_labels = jax.device_put(proto_labels, jax.devices()[0])
|
|
801
|
+
|
|
752
802
|
self.prototypes_ = params['prototypes']
|
|
753
803
|
self.prototype_labels_ = proto_labels
|
|
754
804
|
self.loss_ = float(loss_history[-1]) if len(loss_history) > 0 else None
|
|
@@ -799,6 +849,18 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
799
849
|
if X.ndim != 2:
|
|
800
850
|
raise ValueError(f"X must be 2D, got shape {X.shape}")
|
|
801
851
|
|
|
852
|
+
# Multi-device sharding setup
|
|
853
|
+
if self.devices is not None:
|
|
854
|
+
from prosemble.core.distributed import create_mesh, shard_data
|
|
855
|
+
self._mesh = create_mesh(self.devices)
|
|
856
|
+
n_devices = len(self.devices)
|
|
857
|
+
if self.batch_size is not None and self.batch_size % n_devices != 0:
|
|
858
|
+
raise ValueError(
|
|
859
|
+
f"batch_size ({self.batch_size}) must be divisible by "
|
|
860
|
+
f"number of devices ({n_devices}) for data parallelism."
|
|
861
|
+
)
|
|
862
|
+
X, y = shard_data(X, y, self._mesh)
|
|
863
|
+
|
|
802
864
|
self.classes_ = jnp.unique(y)
|
|
803
865
|
self.n_classes_ = int(len(self.classes_))
|
|
804
866
|
|
|
@@ -855,6 +917,12 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
855
917
|
else:
|
|
856
918
|
opt_state = self._active_optimizer.init(params)
|
|
857
919
|
|
|
920
|
+
# Replicate params and optimizer state across devices
|
|
921
|
+
if self._mesh is not None:
|
|
922
|
+
from prosemble.core.distributed import replicate_params, replicate_opt_state
|
|
923
|
+
params = replicate_params(params, self._mesh)
|
|
924
|
+
opt_state = replicate_opt_state(opt_state, self._mesh)
|
|
925
|
+
|
|
858
926
|
if self._callbacks:
|
|
859
927
|
return self._fit_with_callbacks(X, y, params, opt_state, proto_labels)
|
|
860
928
|
else:
|
|
@@ -881,6 +949,11 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
881
949
|
X = jnp.asarray(X, dtype=jnp.float32)
|
|
882
950
|
y = jnp.asarray(y, dtype=jnp.int32)
|
|
883
951
|
|
|
952
|
+
# Shard data if multi-device
|
|
953
|
+
if self._mesh is not None:
|
|
954
|
+
from prosemble.core.distributed import shard_data, replicate_params, replicate_opt_state
|
|
955
|
+
X, y = shard_data(X, y, self._mesh)
|
|
956
|
+
|
|
884
957
|
# Build params from fitted state
|
|
885
958
|
params = self._get_resume_params({'prototypes': self.prototypes_})
|
|
886
959
|
proto_labels = self.prototype_labels_
|
|
@@ -891,6 +964,12 @@ class SupervisedPrototypeModel(SerializationMixin, MetadataCollectorMixin, Quant
|
|
|
891
964
|
if not hasattr(self, '_opt_state') or self._opt_state is None:
|
|
892
965
|
self._opt_state = self._active_optimizer.init(params)
|
|
893
966
|
|
|
967
|
+
# Replicate params, optimizer state, and proto_labels for multi-device
|
|
968
|
+
if self._mesh is not None:
|
|
969
|
+
params = replicate_params(params, self._mesh)
|
|
970
|
+
self._opt_state = replicate_opt_state(self._opt_state, self._mesh)
|
|
971
|
+
proto_labels = replicate_params({'_': proto_labels}, self._mesh)['_']
|
|
972
|
+
|
|
894
973
|
# Single gradient step
|
|
895
974
|
loss, grads = self._value_and_grad_fn(params, X, y, proto_labels)
|
|
896
975
|
updates, self._opt_state = self._active_optimizer.update(
|
|
@@ -95,6 +95,7 @@ prosemble/core/callbacks.py
|
|
|
95
95
|
prosemble/core/competitions.py
|
|
96
96
|
prosemble/core/data.py
|
|
97
97
|
prosemble/core/distance.py
|
|
98
|
+
prosemble/core/distributed.py
|
|
98
99
|
prosemble/core/initializers.py
|
|
99
100
|
prosemble/core/kernel.py
|
|
100
101
|
prosemble/core/losses.py
|
|
@@ -221,6 +222,7 @@ tests/test_data.py
|
|
|
221
222
|
tests/test_dataset.py
|
|
222
223
|
tests/test_dataset_jax.py
|
|
223
224
|
tests/test_distance_jax.py
|
|
225
|
+
tests/test_distributed.py
|
|
224
226
|
tests/test_examples.sh
|
|
225
227
|
tests/test_fcm_jax.py
|
|
226
228
|
tests/test_glvq.py
|
|
@@ -10,7 +10,7 @@ sys.path.insert(0, os.path.abspath('..'))
|
|
|
10
10
|
project = 'Prosemble'
|
|
11
11
|
copyright = '2024-2026, Nana Abeka Otoo'
|
|
12
12
|
author = 'Nana Abeka Otoo'
|
|
13
|
-
release = '2.
|
|
13
|
+
release = '2.1.1'
|
|
14
14
|
|
|
15
15
|
# -- General configuration ---------------------------------------------------
|
|
16
16
|
extensions = [
|
|
@@ -621,6 +621,128 @@ JAX automatically uses available GPUs. Check your device:
|
|
|
621
621
|
All prosemble operations (training, inference, distance computation) run
|
|
622
622
|
on whatever device JAX is configured to use. No code changes needed.
|
|
623
623
|
|
|
624
|
+
Multi-Device Data Parallelism
|
|
625
|
+
------------------------------
|
|
626
|
+
|
|
627
|
+
For large datasets or image models with trainable backbones, prosemble
|
|
628
|
+
supports data-parallel training across multiple GPUs or TPUs using JAX's
|
|
629
|
+
modern sharding API.
|
|
630
|
+
|
|
631
|
+
.. code-block:: python
|
|
632
|
+
|
|
633
|
+
import jax
|
|
634
|
+
from prosemble.models import GLVQ
|
|
635
|
+
|
|
636
|
+
# Use all available devices
|
|
637
|
+
devices = jax.devices() # e.g., [GpuDevice(id=0), GpuDevice(id=1), ...]
|
|
638
|
+
|
|
639
|
+
model = GLVQ(
|
|
640
|
+
n_prototypes_per_class=2,
|
|
641
|
+
max_iter=100,
|
|
642
|
+
lr=0.01,
|
|
643
|
+
devices=devices,
|
|
644
|
+
use_scan=False,
|
|
645
|
+
)
|
|
646
|
+
model.fit(X_train, y_train)
|
|
647
|
+
|
|
648
|
+
When ``devices`` is specified:
|
|
649
|
+
|
|
650
|
+
- Training data is **sharded** along the batch dimension across devices
|
|
651
|
+
- Model parameters and optimizer state are **replicated** on all devices
|
|
652
|
+
- Gradients are automatically aggregated (all-reduced) across devices
|
|
653
|
+
- After training, parameters are moved back to a single device for inference
|
|
654
|
+
|
|
655
|
+
**Requirements:**
|
|
656
|
+
|
|
657
|
+
- ``batch_size`` (if set) must be divisible by the number of devices
|
|
658
|
+
- The number of training samples should be divisible by the number of devices
|
|
659
|
+
- ``use_scan=False`` is recommended for multi-device training
|
|
660
|
+
|
|
661
|
+
**Primary use case:** Image variants (``ImageGLVQ``, ``ImageGMLVQ``,
|
|
662
|
+
``ImageGTLVQ``, ``ImageCBC``) with trainable CNN backbones on large datasets
|
|
663
|
+
across multiple GPUs/TPUs.
|
|
664
|
+
|
|
665
|
+
.. code-block:: python
|
|
666
|
+
|
|
667
|
+
from prosemble.models import ImageGLVQ
|
|
668
|
+
|
|
669
|
+
model = ImageGLVQ(
|
|
670
|
+
n_prototypes_per_class=1,
|
|
671
|
+
max_iter=50,
|
|
672
|
+
lr=0.001,
|
|
673
|
+
devices=jax.devices(),
|
|
674
|
+
gradient_checkpointing=True, # save memory for deep backbones
|
|
675
|
+
use_scan=False,
|
|
676
|
+
)
|
|
677
|
+
model.fit(X_images, y_labels)
|
|
678
|
+
|
|
679
|
+
``partial_fit()`` also works with multi-device models:
|
|
680
|
+
|
|
681
|
+
.. code-block:: python
|
|
682
|
+
|
|
683
|
+
model.fit(X_batch1, y_batch1)
|
|
684
|
+
model.partial_fit(X_batch2, y_batch2) # data is automatically sharded
|
|
685
|
+
|
|
686
|
+
Gradient Checkpointing
|
|
687
|
+
-----------------------
|
|
688
|
+
|
|
689
|
+
Gradient checkpointing (``jax.remat``) trades compute for memory by
|
|
690
|
+
recomputing forward activations during the backward pass instead of
|
|
691
|
+
storing them. This is beneficial when training models with deep backbones
|
|
692
|
+
(Image, Siamese variants) that would otherwise exhaust GPU memory.
|
|
693
|
+
|
|
694
|
+
.. code-block:: python
|
|
695
|
+
|
|
696
|
+
from prosemble.models import GLVQ
|
|
697
|
+
|
|
698
|
+
model = GLVQ(
|
|
699
|
+
n_prototypes_per_class=2,
|
|
700
|
+
max_iter=100,
|
|
701
|
+
lr=0.01,
|
|
702
|
+
gradient_checkpointing=True,
|
|
703
|
+
)
|
|
704
|
+
model.fit(X_train, y_train)
|
|
705
|
+
|
|
706
|
+
When ``gradient_checkpointing=True``:
|
|
707
|
+
|
|
708
|
+
- Forward activations are **not** stored during the forward pass
|
|
709
|
+
- During backpropagation, the forward pass is **recomputed** to obtain
|
|
710
|
+
activations needed for gradient computation
|
|
711
|
+
- Memory usage is reduced at the cost of ~33% more compute
|
|
712
|
+
- Results are numerically identical to standard training
|
|
713
|
+
|
|
714
|
+
This option has no effect on models with shallow loss functions (e.g.,
|
|
715
|
+
standard GLVQ with Euclidean distance). It provides significant memory
|
|
716
|
+
savings when unfreezing deep CNN backbones in Image or Siamese variants.
|
|
717
|
+
|
|
718
|
+
Custom Gradient Rules
|
|
719
|
+
----------------------
|
|
720
|
+
|
|
721
|
+
Prosemble provides infrastructure for subclasses to define custom
|
|
722
|
+
backward passes using ``jax.custom_vjp``. This is useful for:
|
|
723
|
+
|
|
724
|
+
- Optimal transport distances with non-differentiable sorting operations
|
|
725
|
+
- Hard thresholding operations that need surrogate gradients
|
|
726
|
+
- Distances with numerically unstable standard autodiff
|
|
727
|
+
|
|
728
|
+
Subclasses opt in by setting ``self._use_custom_vjp = True`` in their
|
|
729
|
+
``__init__`` and overriding the ``_custom_vjp_loss`` method:
|
|
730
|
+
|
|
731
|
+
.. code-block:: python
|
|
732
|
+
|
|
733
|
+
class MyModel(GLVQ):
|
|
734
|
+
def __init__(self, **kwargs):
|
|
735
|
+
super().__init__(**kwargs)
|
|
736
|
+
self._use_custom_vjp = True
|
|
737
|
+
|
|
738
|
+
def _custom_vjp_loss(self, params, X, y, proto_labels):
|
|
739
|
+
# Implement custom forward + backward pass
|
|
740
|
+
...
|
|
741
|
+
|
|
742
|
+
When ``_use_custom_vjp`` is active, gradient checkpointing is
|
|
743
|
+
automatically disabled (custom VJP already controls what gets
|
|
744
|
+
saved/recomputed).
|
|
745
|
+
|
|
624
746
|
Training Callbacks
|
|
625
747
|
------------------
|
|
626
748
|
|