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.
Files changed (279) hide show
  1. {prosemble-2.1.0 → prosemble-2.1.1}/.github/workflows/release.yml +1 -1
  2. {prosemble-2.1.0 → prosemble-2.1.1}/PKG-INFO +1 -1
  3. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/__init__.py +3 -0
  4. prosemble-2.1.1/prosemble/core/distributed.py +104 -0
  5. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/glvq.py +3 -1
  6. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lvq1.py +3 -1
  7. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lvq21.py +3 -1
  8. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/median_lvq.py +3 -1
  9. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/probabilistic_lvq.py +6 -2
  10. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/prototype_base.py +89 -10
  11. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/PKG-INFO +1 -1
  12. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/SOURCES.txt +2 -0
  13. {prosemble-2.1.0 → prosemble-2.1.1}/pyproject.toml +1 -1
  14. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/api/core.rst +7 -0
  15. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/conf.py +1 -1
  16. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/advanced.rst +122 -0
  17. prosemble-2.1.1/tests/test_distributed.py +249 -0
  18. {prosemble-2.1.0 → prosemble-2.1.1}/uv.lock +1 -1
  19. {prosemble-2.1.0 → prosemble-2.1.1}/.bumpversion.cfg +0 -0
  20. {prosemble-2.1.0 → prosemble-2.1.1}/.editorconfig +0 -0
  21. {prosemble-2.1.0 → prosemble-2.1.1}/.envrc +0 -0
  22. {prosemble-2.1.0 → prosemble-2.1.1}/.github/ISSUE_TEMPLATE/bug_report.md +0 -0
  23. {prosemble-2.1.0 → prosemble-2.1.1}/.github/ISSUE_TEMPLATE/config.yml +0 -0
  24. {prosemble-2.1.0 → prosemble-2.1.1}/.github/ISSUE_TEMPLATE/feature_request.md +0 -0
  25. {prosemble-2.1.0 → prosemble-2.1.1}/.github/workflows/ci.yml +0 -0
  26. {prosemble-2.1.0 → prosemble-2.1.1}/.gitignore +0 -0
  27. {prosemble-2.1.0 → prosemble-2.1.1}/.readthedocs.yaml +0 -0
  28. {prosemble-2.1.0 → prosemble-2.1.1}/CONSTITUTION.md +0 -0
  29. {prosemble-2.1.0 → prosemble-2.1.1}/LICENSE +0 -0
  30. {prosemble-2.1.0 → prosemble-2.1.1}/MANIFEST.in +0 -0
  31. {prosemble-2.1.0 → prosemble-2.1.1}/ONNX_COVERAGE.md +0 -0
  32. {prosemble-2.1.0 → prosemble-2.1.1}/README.md +0 -0
  33. {prosemble-2.1.0 → prosemble-2.1.1}/TASK.md +0 -0
  34. {prosemble-2.1.0 → prosemble-2.1.1}/benchmarks/benchmark_distance_jax.py +0 -0
  35. {prosemble-2.1.0 → prosemble-2.1.1}/benchmarks/benchmark_fcm_jax.py +0 -0
  36. {prosemble-2.1.0 → prosemble-2.1.1}/benchmarks/benchmark_pcm_jax.py +0 -0
  37. {prosemble-2.1.0 → prosemble-2.1.1}/devenv.lock +0 -0
  38. {prosemble-2.1.0 → prosemble-2.1.1}/devenv.nix +0 -0
  39. {prosemble-2.1.0 → prosemble-2.1.1}/devenv.yaml +0 -0
  40. {prosemble-2.1.0 → prosemble-2.1.1}/examples/afcm_iris.py +0 -0
  41. {prosemble-2.1.0 → prosemble-2.1.1}/examples/bgpc_iris.py +0 -0
  42. {prosemble-2.1.0 → prosemble-2.1.1}/examples/cbc_iris.py +0 -0
  43. {prosemble-2.1.0 → prosemble-2.1.1}/examples/celvq_ng_iris.py +0 -0
  44. {prosemble-2.1.0 → prosemble-2.1.1}/examples/fcm_iris.py +0 -0
  45. {prosemble-2.1.0 → prosemble-2.1.1}/examples/fpcm_iris.py +0 -0
  46. {prosemble-2.1.0 → prosemble-2.1.1}/examples/glvq_advanced_iris.py +0 -0
  47. {prosemble-2.1.0 → prosemble-2.1.1}/examples/glvq_iris.py +0 -0
  48. {prosemble-2.1.0 → prosemble-2.1.1}/examples/glvq_unequal_prototypes_iris.py +0 -0
  49. {prosemble-2.1.0 → prosemble-2.1.1}/examples/gmlvq_iris.py +0 -0
  50. {prosemble-2.1.0 → prosemble-2.1.1}/examples/grlvq_iris.py +0 -0
  51. {prosemble-2.1.0 → prosemble-2.1.1}/examples/hcm_iris.py +0 -0
  52. {prosemble-2.1.0 → prosemble-2.1.1}/examples/heskes_som_iris.py +0 -0
  53. {prosemble-2.1.0 → prosemble-2.1.1}/examples/ipcm2_iris.py +0 -0
  54. {prosemble-2.1.0 → prosemble-2.1.1}/examples/ipcm_iris.py +0 -0
  55. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kafcm_iris.py +0 -0
  56. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kfcm_iris.py +0 -0
  57. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kfpcm_iris.py +0 -0
  58. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kipcm2_iris.py +0 -0
  59. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kipcm_iris.py +0 -0
  60. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kmeans_iris.py +0 -0
  61. {prosemble-2.1.0 → prosemble-2.1.1}/examples/knn_iris.py +0 -0
  62. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kohonen_som_iris.py +0 -0
  63. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kpcm_iris.py +0 -0
  64. {prosemble-2.1.0 → prosemble-2.1.1}/examples/kpfcm_iris.py +0 -0
  65. {prosemble-2.1.0 → prosemble-2.1.1}/examples/large_dataset_training.py +0 -0
  66. {prosemble-2.1.0 → prosemble-2.1.1}/examples/lcelvq_ng_iris.py +0 -0
  67. {prosemble-2.1.0 → prosemble-2.1.1}/examples/lmrslvq_iris.py +0 -0
  68. {prosemble-2.1.0 → prosemble-2.1.1}/examples/lmrslvq_ng_iris.py +0 -0
  69. {prosemble-2.1.0 → prosemble-2.1.1}/examples/lvq1_iris.py +0 -0
  70. {prosemble-2.1.0 → prosemble-2.1.1}/examples/lvqmln_iris.py +0 -0
  71. {prosemble-2.1.0 → prosemble-2.1.1}/examples/mcelvq_ng_iris.py +0 -0
  72. {prosemble-2.1.0 → prosemble-2.1.1}/examples/mrslvq_iris.py +0 -0
  73. {prosemble-2.1.0 → prosemble-2.1.1}/examples/mrslvq_ng_iris.py +0 -0
  74. {prosemble-2.1.0 → prosemble-2.1.1}/examples/neural_gas_iris.py +0 -0
  75. {prosemble-2.1.0 → prosemble-2.1.1}/examples/npc_iris.py +0 -0
  76. {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_lmrslvq_iris.py +0 -0
  77. {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_lmrslvq_ng_iris.py +0 -0
  78. {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_mrslvq_iris.py +0 -0
  79. {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_mrslvq_ng_iris.py +0 -0
  80. {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_rslvq_iris.py +0 -0
  81. {prosemble-2.1.0 → prosemble-2.1.1}/examples/oc_rslvq_ng_iris.py +0 -0
  82. {prosemble-2.1.0 → prosemble-2.1.1}/examples/pcm_iris.py +0 -0
  83. {prosemble-2.1.0 → prosemble-2.1.1}/examples/pfcm_iris.py +0 -0
  84. {prosemble-2.1.0 → prosemble-2.1.1}/examples/plvq_iris.py +0 -0
  85. {prosemble-2.1.0 → prosemble-2.1.1}/examples/rslvq_ng_iris.py +0 -0
  86. {prosemble-2.1.0 → prosemble-2.1.1}/examples/slng_iris.py +0 -0
  87. {prosemble-2.1.0 → prosemble-2.1.1}/examples/smng_iris.py +0 -0
  88. {prosemble-2.1.0 → prosemble-2.1.1}/examples/som_iris.py +0 -0
  89. {prosemble-2.1.0 → prosemble-2.1.1}/examples/srng_iris.py +0 -0
  90. {prosemble-2.1.0 → prosemble-2.1.1}/examples/stng_iris.py +0 -0
  91. {prosemble-2.1.0 → prosemble-2.1.1}/examples/tcelvq_ng_iris.py +0 -0
  92. {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_fcm_iris.py +0 -0
  93. {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_glvq_iris.py +0 -0
  94. {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_neural_gas_iris.py +0 -0
  95. {prosemble-2.1.0 → prosemble-2.1.1}/examples/vis_som_iris.py +0 -0
  96. {prosemble-2.1.0 → prosemble-2.1.1}/flake.nix +0 -0
  97. {prosemble-2.1.0 → prosemble-2.1.1}/mkdocs.yml +0 -0
  98. {prosemble-2.1.0 → prosemble-2.1.1}/nix/package.nix +0 -0
  99. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/__init__.py +0 -0
  100. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/activations.py +0 -0
  101. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/callbacks.py +0 -0
  102. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/competitions.py +0 -0
  103. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/data.py +0 -0
  104. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/distance.py +0 -0
  105. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/initializers.py +0 -0
  106. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/kernel.py +0 -0
  107. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/losses.py +0 -0
  108. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/manifolds.py +0 -0
  109. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/model_selection.py +0 -0
  110. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/onnx_export.py +0 -0
  111. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/pipeline.py +0 -0
  112. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/pooling.py +0 -0
  113. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/protocols.py +0 -0
  114. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/quantization.py +0 -0
  115. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/serialization.py +0 -0
  116. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/similarities.py +0 -0
  117. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/utils.py +0 -0
  118. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/vis.py +0 -0
  119. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/core/visualizer.py +0 -0
  120. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/datasets/__init__.py +0 -0
  121. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/datasets/dataset.py +0 -0
  122. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/__init__.py +0 -0
  123. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/afcm.py +0 -0
  124. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/base.py +0 -0
  125. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/bgpc.py +0 -0
  126. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/cbc.py +0 -0
  127. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/celvq_ng.py +0 -0
  128. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/celvq_ng_mixin.py +0 -0
  129. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/crossentropy_lvq.py +0 -0
  130. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/fcm.py +0 -0
  131. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/fpcm.py +0 -0
  132. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/growing_neural_gas.py +0 -0
  133. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/hcm.py +0 -0
  134. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/heskes_som.py +0 -0
  135. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/image_cbc.py +0 -0
  136. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/image_lvq.py +0 -0
  137. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/ipcm.py +0 -0
  138. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/ipcm2.py +0 -0
  139. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kafcm.py +0 -0
  140. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kfcm.py +0 -0
  141. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kfpcm.py +0 -0
  142. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kipcm.py +0 -0
  143. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kipcm2.py +0 -0
  144. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kmeans.py +0 -0
  145. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/knn.py +0 -0
  146. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kohonen_som.py +0 -0
  147. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kpcm.py +0 -0
  148. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/kpfcm.py +0 -0
  149. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lcelvq_ng.py +0 -0
  150. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/local_matrix_lvq.py +0 -0
  151. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/lvqmln.py +0 -0
  152. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/matrix_lvq.py +0 -0
  153. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/matrix_rslvq.py +0 -0
  154. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/mcelvq_ng.py +0 -0
  155. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/mrslvq_ng.py +0 -0
  156. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/neural_gas.py +0 -0
  157. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/npc.py +0 -0
  158. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_glvq.py +0 -0
  159. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_glvq_ng.py +0 -0
  160. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_glvq_ng_mixin.py +0 -0
  161. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gmlvq.py +0 -0
  162. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gmlvq_ng.py +0 -0
  163. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_grlvq.py +0 -0
  164. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_grlvq_ng.py +0 -0
  165. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gtlvq.py +0 -0
  166. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_gtlvq_ng.py +0 -0
  167. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_lgmlvq.py +0 -0
  168. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_lgmlvq_ng.py +0 -0
  169. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_mrslvq.py +0 -0
  170. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_mrslvq_ng.py +0 -0
  171. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_rslvq.py +0 -0
  172. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_rslvq_ng.py +0 -0
  173. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/oc_rslvq_ng_mixin.py +0 -0
  174. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/pcm.py +0 -0
  175. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/pfcm.py +0 -0
  176. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/plvq.py +0 -0
  177. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/relevance_lvq.py +0 -0
  178. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_neural_gas.py +0 -0
  179. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_slng.py +0 -0
  180. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_smng.py +0 -0
  181. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_srng.py +0 -0
  182. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/riemannian_stng.py +0 -0
  183. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/rslvq_ng.py +0 -0
  184. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/siamese_lvq.py +0 -0
  185. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/slng.py +0 -0
  186. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/smng.py +0 -0
  187. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/som.py +0 -0
  188. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/srng.py +0 -0
  189. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/stng.py +0 -0
  190. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ.py +0 -0
  191. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_lm.py +0 -0
  192. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_m.py +0 -0
  193. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_r.py +0 -0
  194. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/svq_occ_t.py +0 -0
  195. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/tangent_lvq.py +0 -0
  196. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/models/tcelvq_ng.py +0 -0
  197. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble/py.typed +0 -0
  198. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/dependency_links.txt +0 -0
  199. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/requires.txt +0 -0
  200. {prosemble-2.1.0 → prosemble-2.1.1}/prosemble.egg-info/top_level.txt +0 -0
  201. {prosemble-2.1.0 → prosemble-2.1.1}/requirements_docs.txt +0 -0
  202. {prosemble-2.1.0 → prosemble-2.1.1}/setup.cfg +0 -0
  203. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/Makefile +0 -0
  204. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/api/models.rst +0 -0
  205. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/clustering.rst +0 -0
  206. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/one_class.rst +0 -0
  207. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/onnx.rst +0 -0
  208. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/riemannian.rst +0 -0
  209. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/supervised.rst +0 -0
  210. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/guides/topology.rst +0 -0
  211. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/index.rst +0 -0
  212. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/installation.rst +0 -0
  213. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/make.bat +0 -0
  214. {prosemble-2.1.0 → prosemble-2.1.1}/sphinx-docs/quickstart.rst +0 -0
  215. {prosemble-2.1.0 → prosemble-2.1.1}/tests/__init__.py +0 -0
  216. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_activations.py +0 -0
  217. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_cbc.py +0 -0
  218. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_celvq_ng.py +0 -0
  219. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_checkpoint.py +0 -0
  220. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_clustering_patience.py +0 -0
  221. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_competitions.py +0 -0
  222. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_core.py +0 -0
  223. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_core_jax.py +0 -0
  224. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_data.py +0 -0
  225. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_dataset.py +0 -0
  226. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_dataset_jax.py +0 -0
  227. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_distance_jax.py +0 -0
  228. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_examples.sh +0 -0
  229. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_fcm_jax.py +0 -0
  230. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_glvq.py +0 -0
  231. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_gmlvq.py +0 -0
  232. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_heskes_som.py +0 -0
  233. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_image_lvq.py +0 -0
  234. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_initializers.py +0 -0
  235. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_jax_models.py +0 -0
  236. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_jit_inference.py +0 -0
  237. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_kohonen_som.py +0 -0
  238. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_lcelvq_ng.py +0 -0
  239. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_losses.py +0 -0
  240. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_lvq_nongradient.py +0 -0
  241. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_lvqmln_plvq.py +0 -0
  242. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_matrix_rslvq.py +0 -0
  243. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_mcelvq_ng.py +0 -0
  244. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_minibatch.py +0 -0
  245. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_mixed_precision.py +0 -0
  246. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_model_selection.py +0 -0
  247. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_mrslvq_ng.py +0 -0
  248. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_neural_gas.py +0 -0
  249. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_glvq.py +0 -0
  250. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_glvq_ng.py +0 -0
  251. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_glvq_variants.py +0 -0
  252. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_mrslvq.py +0 -0
  253. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_rslvq.py +0 -0
  254. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_oc_rslvq_ng.py +0 -0
  255. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_onnx_export.py +0 -0
  256. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pcm_jax.py +0 -0
  257. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pfcm_jax.py +0 -0
  258. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pipeline.py +0 -0
  259. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_pooling.py +0 -0
  260. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_probabilistic.py +0 -0
  261. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_protocols.py +0 -0
  262. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_prototype_base.py +0 -0
  263. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_quantization.py +0 -0
  264. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_riemannian_neural_gas.py +0 -0
  265. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_riemannian_supervised_ng.py +0 -0
  266. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_rslvq_ng.py +0 -0
  267. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_scan_training.py +0 -0
  268. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_serialization.py +0 -0
  269. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_siamese_lvq.py +0 -0
  270. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_similarities.py +0 -0
  271. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_slng.py +0 -0
  272. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_smng.py +0 -0
  273. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_srng.py +0 -0
  274. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_stng.py +0 -0
  275. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_svq_occ.py +0 -0
  276. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_svq_occ_r.py +0 -0
  277. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_svq_occ_variants.py +0 -0
  278. {prosemble-2.1.0 → prosemble-2.1.1}/tests/test_tcelvq_ng.py +0 -0
  279. {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
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: prosemble
3
- Version: 2.1.0
3
+ Version: 2.1.1
4
4
  Summary: A python project for prototype-based machine learning models
5
5
  Author-email: Nana Abeka Otoo <nanaabekaotoo@gmail.com>
6
6
  License: MIT
@@ -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, with optional mixed precision."""
657
- if self._mp_dtype is not None:
658
- scaled_loss, grads = jax.value_and_grad(self._compute_loss_mp)(
659
- params, X, y, proto_labels
660
- )
661
- loss = scaled_loss / self._loss_scale
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
- return loss, grads
664
- return jax.value_and_grad(self._compute_loss)(
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(
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: prosemble
3
- Version: 2.1.0
3
+ Version: 2.1.1
4
4
  Summary: A python project for prototype-based machine learning models
5
5
  Author-email: Nana Abeka Otoo <nanaabekaotoo@gmail.com>
6
6
  License: MIT
@@ -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
@@ -1,6 +1,6 @@
1
1
  [project]
2
2
  name = "prosemble"
3
- version = "2.1.0"
3
+ version = "2.1.1"
4
4
  description = "A python project for prototype-based machine learning models"
5
5
  readme = "README.md"
6
6
  requires-python = ">=3.11"
@@ -141,6 +141,13 @@ Protocols
141
141
  :members:
142
142
  :undoc-members:
143
143
 
144
+ Distributed Training
145
+ --------------------
146
+
147
+ .. automodule:: prosemble.core.distributed
148
+ :members:
149
+ :undoc-members:
150
+
144
151
  Data Utilities
145
152
  --------------
146
153
 
@@ -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 = '2.0.0'
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