braindecode 1.8.1.dev169289523__tar.gz → 1.8.1.dev169440825__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 (209) hide show
  1. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/NOTICE.txt +1 -0
  2. {braindecode-1.8.1.dev169289523/braindecode.egg-info → braindecode-1.8.1.dev169440825}/PKG-INFO +1 -1
  3. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/__init__.py +2 -0
  4. braindecode-1.8.1.dev169440825/braindecode/models/epint.py +311 -0
  5. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/luna.py +4 -3
  6. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/mirepnet.py +4 -1
  7. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/msvtnet.py +115 -0
  8. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/neurorvq.py +3 -2
  9. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/summary.csv +1 -0
  10. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/tfm_tokenizer.py +44 -3
  11. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/util.py +2 -0
  12. braindecode-1.8.1.dev169440825/braindecode/version.py +1 -0
  13. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825/braindecode.egg-info}/PKG-INFO +1 -1
  14. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode.egg-info/SOURCES.txt +1 -0
  15. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/api.rst +1 -0
  16. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/whats_new.rst +11 -0
  17. braindecode-1.8.1.dev169289523/braindecode/version.py +0 -1
  18. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/LICENSE.txt +0 -0
  19. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/MANIFEST.in +0 -0
  20. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/README.rst +0 -0
  21. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/__init__.py +0 -0
  22. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/augmentation/__init__.py +0 -0
  23. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/augmentation/base.py +0 -0
  24. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/augmentation/functional.py +0 -0
  25. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/augmentation/transforms.py +0 -0
  26. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/classifier.py +0 -0
  27. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/__init__.py +0 -0
  28. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/_notebook_viewer.py +0 -0
  29. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/base.py +0 -0
  30. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bbci.py +0 -0
  31. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bcicomp.py +0 -0
  32. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/__init__.py +0 -0
  33. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/datasets.py +0 -0
  34. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/format.py +0 -0
  35. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/hub.py +0 -0
  36. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/hub_format.py +0 -0
  37. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/hub_io.py +0 -0
  38. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/hub_validation.py +0 -0
  39. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/bids/iterable.py +0 -0
  40. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/chb_mit.py +0 -0
  41. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/collate.py +0 -0
  42. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/mne.py +0 -0
  43. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/moabb.py +0 -0
  44. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/nmt.py +0 -0
  45. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/registry.py +0 -0
  46. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/siena.py +0 -0
  47. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/sleep_physio_challe_18.py +0 -0
  48. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/sleep_physionet.py +0 -0
  49. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/tuh.py +0 -0
  50. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/utils.py +0 -0
  51. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datasets/xy.py +0 -0
  52. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datautil/__init__.py +0 -0
  53. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datautil/channel_utils.py +0 -0
  54. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datautil/hub_formats.py +0 -0
  55. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datautil/serialization.py +0 -0
  56. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/datautil/util.py +0 -0
  57. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/eegneuralnet.py +0 -0
  58. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/functional/__init__.py +0 -0
  59. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/functional/_real_dft.py +0 -0
  60. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/functional/functions.py +0 -0
  61. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/functional/initialization.py +0 -0
  62. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/atcnet.py +0 -0
  63. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/attentionbasenet.py +0 -0
  64. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/attn_sleep.py +0 -0
  65. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/axon.py +0 -0
  66. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/barista.py +0 -0
  67. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/base.py +0 -0
  68. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/bendr.py +0 -0
  69. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/biot.py +0 -0
  70. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/brainbert.py +0 -0
  71. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/brainmodule.py +0 -0
  72. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/brainomni.py +0 -0
  73. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/brant.py +0 -0
  74. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/cbramod.py +0 -0
  75. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/codebrain.py +0 -0
  76. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/config.py +0 -0
  77. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/contrawr.py +0 -0
  78. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/csbrain.py +0 -0
  79. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/ctnet.py +0 -0
  80. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/dance.py +0 -0
  81. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/deep4.py +0 -0
  82. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/deepsleepnet.py +0 -0
  83. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/dgcnn.py +0 -0
  84. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/diver1.py +0 -0
  85. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eeg_clip.py +0 -0
  86. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegconformer.py +0 -0
  87. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegdino.py +0 -0
  88. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eeginception_erp.py +0 -0
  89. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eeginception_mi.py +0 -0
  90. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegitnet.py +0 -0
  91. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegminer.py +0 -0
  92. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegnet.py +0 -0
  93. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegnex.py +0 -0
  94. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegpt.py +0 -0
  95. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegsimpleconv.py +0 -0
  96. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegsym.py +0 -0
  97. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/eegtcnet.py +0 -0
  98. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/emg2qwerty.py +0 -0
  99. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/fbcnet.py +0 -0
  100. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/fblightconvnet.py +0 -0
  101. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/fbmsnet.py +0 -0
  102. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/guetschel2026.py +0 -0
  103. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/hybrid.py +0 -0
  104. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/ifnet.py +0 -0
  105. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/labram.py +0 -0
  106. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/mapa.py +0 -0
  107. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/medformer.py +0 -0
  108. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/meta_neuromotor.py +0 -0
  109. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/mscformer.py +0 -0
  110. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/mvpformer.py +0 -0
  111. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/neuropose.py +0 -0
  112. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/neurorvq_tokenizer.py +0 -0
  113. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/patchedtransformer.py +0 -0
  114. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/popt.py +0 -0
  115. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/reve.py +0 -0
  116. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sccnet.py +0 -0
  117. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/seizure_transformer.py +0 -0
  118. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sensingdynamics.py +0 -0
  119. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/shallow_fbcsp.py +0 -0
  120. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/signal_jepa.py +0 -0
  121. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sinc_shallow.py +0 -0
  122. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sleep_stager_blanco_2020.py +0 -0
  123. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sleep_stager_chambon_2018.py +0 -0
  124. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sleepfm.py +0 -0
  125. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sparcnet.py +0 -0
  126. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/sstdpn.py +0 -0
  127. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/steegformer.py +0 -0
  128. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/syncnet.py +0 -0
  129. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/tcformer.py +0 -0
  130. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/tcn.py +0 -0
  131. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/tidnet.py +0 -0
  132. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/tmsanet.py +0 -0
  133. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/tsinception.py +0 -0
  134. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/usleep.py +0 -0
  135. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/vemg2pose.py +0 -0
  136. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/models/zuna.py +0 -0
  137. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/__init__.py +0 -0
  138. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/activation.py +0 -0
  139. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/attention.py +0 -0
  140. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/blocks.py +0 -0
  141. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/channels.py +0 -0
  142. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/convolution.py +0 -0
  143. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/dance_modules.py +0 -0
  144. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/filter.py +0 -0
  145. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/layers.py +0 -0
  146. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/linear.py +0 -0
  147. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/linear_attention.py +0 -0
  148. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/parametrization.py +0 -0
  149. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/popt_modules.py +0 -0
  150. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/quantization.py +0 -0
  151. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/stats.py +0 -0
  152. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/util.py +0 -0
  153. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/modules/wrapper.py +0 -0
  154. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/preprocessing/__init__.py +0 -0
  155. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/preprocessing/eegprep_preprocess.py +0 -0
  156. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/preprocessing/mne_preprocess.py +0 -0
  157. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/preprocessing/preprocess.py +0 -0
  158. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/preprocessing/util.py +0 -0
  159. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/preprocessing/windowers.py +0 -0
  160. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/regressor.py +0 -0
  161. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/samplers/__init__.py +0 -0
  162. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/samplers/base.py +0 -0
  163. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/samplers/ssl.py +0 -0
  164. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/training/__init__.py +0 -0
  165. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/training/callbacks.py +0 -0
  166. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/training/losses.py +0 -0
  167. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/training/scoring.py +0 -0
  168. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/util.py +0 -0
  169. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/__init__.py +0 -0
  170. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/activations.py +0 -0
  171. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/attribution.py +0 -0
  172. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/confusion_matrices.py +0 -0
  173. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/frequency.py +0 -0
  174. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/metrics.py +0 -0
  175. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/sanity.py +0 -0
  176. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode/visualization/topology.py +0 -0
  177. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode.egg-info/dependency_links.txt +0 -0
  178. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode.egg-info/requires.txt +0 -0
  179. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/braindecode.egg-info/top_level.txt +0 -0
  180. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/Makefile +0 -0
  181. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/_templates/autosummary/class.rst +0 -0
  182. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/_templates/autosummary/class_in_subdir.rst +0 -0
  183. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/_templates/autosummary/function.rst +0 -0
  184. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/_templates/autosummary/function_in_subdir.rst +0 -0
  185. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/agent_docs.rst +0 -0
  186. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/cite.rst +0 -0
  187. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/conf.py +0 -0
  188. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/help.rst +0 -0
  189. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/index.rst +0 -0
  190. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/install/install.rst +0 -0
  191. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/install/install_pip.rst +0 -0
  192. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/install/install_source.rst +0 -0
  193. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/attention.rst +0 -0
  194. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/channel.rst +0 -0
  195. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/convolution.rst +0 -0
  196. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/filterbank.rst +0 -0
  197. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/gnn.rst +0 -0
  198. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/interpretable.rst +0 -0
  199. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/lbm.rst +0 -0
  200. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/recurrent.rst +0 -0
  201. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/categorization/spd.rst +0 -0
  202. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/models.rst +0 -0
  203. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/models_categorization.rst +0 -0
  204. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/models_table.rst +0 -0
  205. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/models/models_visualization.rst +0 -0
  206. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/sg_execution_times.rst +0 -0
  207. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/docs/user_guide/channel_strategies.rst +0 -0
  208. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/pyproject.toml +0 -0
  209. {braindecode-1.8.1.dev169289523 → braindecode-1.8.1.dev169440825}/setup.cfg +0 -0
@@ -54,6 +54,7 @@ The following components are licensed under the MIT License:
54
54
  - `docs/_static/model/seizure_transformer_arch.png`
55
55
  - `braindecode/models/guetschel2026.py`
56
56
  - `braindecode/models/tmsanet.py`
57
+ - `braindecode/models/epint.py`
57
58
 
58
59
  The `brainomni.py` file is a derivative of `OpenTSLab/BrainOmni` (MIT), whose
59
60
  SEANet codec derives from Meta's `facebookresearch/encodec` (MIT). The
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: braindecode
3
- Version: 1.8.1.dev169289523
3
+ Version: 1.8.1.dev169440825
4
4
  Summary: Deep learning software to decode EEG, ECG or MEG signals
5
5
  Author-email: Robin Tibor Schirrmeister <robintibor@gmail.com>, Bruno Aristimunha Pinto <b.aristimunha@gmail.com>, Alexandre Gramfort <agramfort@meta.com>
6
6
  Maintainer-email: Alexandre Gramfort <agramfort@meta.com>, Bruno Aristimunha Pinto <b.aristimunha@gmail.com>, Robin Tibor Schirrmeister <robintibor@gmail.com>
@@ -36,6 +36,7 @@ from .eegsimpleconv import EEGSimpleConv
36
36
  from .eegsym import EEGSym
37
37
  from .eegtcnet import EEGTCNet
38
38
  from .emg2qwerty import EMG2QwertyNet
39
+ from .epint import EpiNT
39
40
  from .fbcnet import FBCNet
40
41
  from .fblightconvnet import FBLightConvNet
41
42
  from .fbmsnet import FBMSNet
@@ -134,6 +135,7 @@ __all__ = [
134
135
  "EEGTCNet",
135
136
  "DGCNN",
136
137
  "EMG2QwertyNet",
138
+ "EpiNT",
137
139
  "NeuroPose",
138
140
  "SensingDynamics",
139
141
  "VEMG2Pose",
@@ -0,0 +1,311 @@
1
+ # Authors: Runkai Zhang <271013216@qq.com>
2
+ # Julien Gadonneix (braindecode adaptation)
3
+ #
4
+ # License: MIT
5
+ # Adapted from https://github.com/RunKZhang/EpiNT (MIT).
6
+
7
+ from __future__ import annotations
8
+
9
+ import torch
10
+ import torch.nn as nn
11
+ import torch.nn.functional as F
12
+ from einops.layers.torch import Rearrange, Reduce
13
+
14
+ from braindecode.functional import rotate_pairs
15
+ from braindecode.models.base import EEGModuleMixin
16
+ from braindecode.modules import FeedForwardBlock
17
+
18
+
19
+ class EpiNT(EEGModuleMixin, nn.Module, license="mit"):
20
+ r"""EpiNT from Zhang et al. (2025) [EpiNT2025]_.
21
+
22
+ :bdg-danger:`Foundation Model` :bdg-info:`Attention/Transformer`
23
+
24
+ .. figure:: https://raw.githubusercontent.com/RunKZhang/EpiNT/master/assets/ModelStructure.png
25
+ :align: center
26
+ :alt: EpiNT architecture
27
+
28
+ .. versionadded:: 1.9
29
+
30
+ EpiNT (Epilepsy Neurophysiological Transformer) is pretrained on 2,700 hours
31
+ of scalp EEG and intracranial EEG (ECoG, sEEG) from 1,199 patients. It is
32
+ **channel independent**: every channel is encoded alone, so one checkpoint
33
+ serves any montage and modality.
34
+
35
+ .. rubric:: Architecture Overview
36
+
37
+ 1. ``instance_norm`` z-scores each channel over time (affine, one scale and
38
+ one offset shared by all channels).
39
+ 2. ``patch_embedding`` projects non-overlapping patches of ``patch_size``
40
+ samples to ``embed_dim``, and a learned ``cls_token`` is prepended.
41
+ 3. ``encoder`` is a stack of post-norm Transformer layers with rotary
42
+ position embeddings (RoPE) in the self-attention. As upstream, each
43
+ ``LayerNorm`` spans the token **and** feature axes, so the number of
44
+ patches (``n_times // patch_size``) is fixed at construction.
45
+ 4. The class token of each channel is averaged over channels and
46
+ ``final_layer`` (dropout and one linear layer) classifies it.
47
+
48
+ The pretraining mask token and frequency-domain mapping quantizer are not
49
+ used downstream, so they are not part of this module.
50
+
51
+ Parameters
52
+ ----------
53
+ patch_size : int, optional
54
+ Samples per patch. Default 256.
55
+ embed_dim : int, optional
56
+ Transformer width. Default 512.
57
+ n_layers : int, optional
58
+ Number of Transformer layers. Default 6.
59
+ n_heads : int, optional
60
+ Number of attention heads; ``embed_dim // n_heads`` must be even for
61
+ RoPE. Default 8.
62
+ ffn_dim : int, optional
63
+ Inner dimension of the feed-forward blocks. Default 2048.
64
+ drop_prob : float, optional
65
+ Dropout in the Transformer layers. Default 0.1.
66
+ head_drop_prob : float, optional
67
+ Dropout before ``final_layer``. Default 0.1.
68
+ rope_theta : float, optional
69
+ RoPE base frequency. Default 10000.
70
+ activation : type[nn.Module], optional
71
+ Feed-forward activation. Default ``nn.ReLU``, as pretrained.
72
+
73
+ Notes
74
+ -----
75
+ **Input.** The pretrained weights expect ``n_times=3072``: 12 s at 256 Hz
76
+ (scalp EEG) or 3 s at 1024 Hz (iEEG), in microvolts as the authors prepare
77
+ them (``instance_norm`` has ``eps=1e-5``, so signals in volts are not scaled
78
+ to unit variance). The authors classify single-channel samples
79
+ (``n_chans=1``).
80
+
81
+ **Weights.** The authors' weights are plain files in their repository, not
82
+ a release (`License <https://github.com/RunKZhang/EpiNT/blob/master/LICENSE>`_):
83
+ ``weights/representations.bin`` (the pretrained backbone) and
84
+ ``weights/finetune_on_MAYO.bin`` (a linear probe of it on MAYO). Both load
85
+ through :attr:`mapping`, the probe with ``strict=True``::
86
+
87
+ state = torch.load("representations.bin", map_location="cpu")
88
+ model = EpiNT(n_chans=1, n_outputs=2, n_times=3072)
89
+ model.load_state_dict(state, strict=False) # no head, mask token unused
90
+
91
+ The authors' fine-tuning code at the time of the paper applies the
92
+ cross-entropy to softmax outputs (changed upstream in August 2025); with it
93
+ and seed 0 their code trains a MAYO probe with the scores of
94
+ ``finetune_on_MAYO.bin`` (F1 0.916, accuracy 0.851).
95
+
96
+ **Replication.** With the released weights, the port matches the authors'
97
+ code within float rounding (logits within 1.4e-6 in float32). Fine-tuned
98
+ with the authors' protocol, the three MAYO cells of Table 3 (the authors'
99
+ test patients) are within 3 % of the paper. Linear probing on CHB-MIT and
100
+ FNUSA gives 0.73-0.76 with both the port and the authors' code (paper: 0.857
101
+ and 0.805); the paper does not publish its test patients for these two
102
+ datasets or the runs behind its mean ± std.
103
+
104
+ References
105
+ ----------
106
+ .. [EpiNT2025] Zhang, R., Yu, H., Gan, J. Q. and Wang, H., 2025.
107
+ Cross-modal epileptic signal harmonization: Frequency domain mapping
108
+ quantization for pre-training a unified neurophysiological transformer.
109
+ arXiv:2506.17068.
110
+ Code: https://github.com/RunKZhang/EpiNT
111
+ """
112
+
113
+ def __init__(
114
+ self,
115
+ patch_size: int = 256,
116
+ embed_dim: int = 512,
117
+ n_layers: int = 6,
118
+ n_heads: int = 8,
119
+ ffn_dim: int = 2048,
120
+ drop_prob: float = 0.1,
121
+ head_drop_prob: float = 0.1,
122
+ rope_theta: float = 10000.0,
123
+ activation: type[nn.Module] = nn.ReLU,
124
+ # braindecode signal parameters
125
+ n_outputs=None,
126
+ n_chans=None,
127
+ chs_info=None,
128
+ n_times=None,
129
+ input_window_seconds=None,
130
+ sfreq=None,
131
+ ):
132
+ super().__init__(
133
+ n_outputs=n_outputs,
134
+ n_chans=n_chans,
135
+ chs_info=chs_info,
136
+ n_times=n_times,
137
+ input_window_seconds=input_window_seconds,
138
+ sfreq=sfreq,
139
+ )
140
+ del n_outputs, n_chans, chs_info, n_times, input_window_seconds, sfreq
141
+
142
+ if embed_dim % n_heads or (embed_dim // n_heads) % 2:
143
+ raise ValueError(
144
+ f"embed_dim={embed_dim} must split into n_heads={n_heads} heads "
145
+ "of even width for RoPE."
146
+ )
147
+ self.patch_size = patch_size
148
+ self.n_patches = self.n_times // patch_size
149
+ if self.n_patches < 1:
150
+ raise ValueError(
151
+ f"n_times={self.n_times} is shorter than one patch "
152
+ f"(patch_size={patch_size})."
153
+ )
154
+ n_tokens = self.n_patches + 1
155
+
156
+ self.merge_channels = Rearrange("batch chans times -> (batch chans) 1 times")
157
+ self.instance_norm = nn.InstanceNorm1d(1, affine=True)
158
+ self.patchify = Rearrange(
159
+ "batch 1 (patches size) -> batch patches size", size=patch_size
160
+ )
161
+ self.patch_embedding = nn.Linear(patch_size, embed_dim)
162
+ self.cls_token = nn.Parameter(torch.empty(embed_dim))
163
+ nn.init.normal_(self.cls_token, std=0.1)
164
+ # Interleaved-pair RoPE angles, as ``precompute_freqs_cis`` upstream.
165
+ head_dim = embed_dim // n_heads
166
+ inv_freq = 1.0 / rope_theta ** (torch.arange(0, head_dim, 2).float() / head_dim)
167
+ angles = torch.outer(torch.arange(n_tokens).float(), inv_freq)
168
+ angles = angles.repeat_interleave(2, dim=-1) # (tokens, head_dim)
169
+ self.register_buffer("rope_cos", angles.cos(), persistent=False)
170
+ self.register_buffer("rope_sin", angles.sin(), persistent=False)
171
+ self.encoder = nn.ModuleList(
172
+ [
173
+ _EpiNTLayer(
174
+ n_tokens, embed_dim, n_heads, ffn_dim, drop_prob, activation
175
+ )
176
+ for _ in range(n_layers)
177
+ ]
178
+ )
179
+ self.split_channels = Rearrange(
180
+ "(batch chans) dim -> batch chans dim", chans=self.n_chans
181
+ )
182
+ self.pool = Reduce("batch chans dim -> batch dim", "mean")
183
+ self.final_layer = nn.Sequential(
184
+ nn.Dropout(head_drop_prob), nn.Linear(embed_dim, self.n_outputs)
185
+ )
186
+ nn.init.kaiming_uniform_(self.final_layer[1].weight)
187
+ nn.init.zeros_(self.final_layer[1].bias)
188
+
189
+ self.mapping = {
190
+ "norm.weight": "instance_norm.weight",
191
+ "norm.bias": "instance_norm.bias",
192
+ "embed.proj.weight": "patch_embedding.weight",
193
+ "embed.proj.bias": "patch_embedding.bias",
194
+ "embed.cls_embed": "cls_token",
195
+ "head.weight": "final_layer.1.weight",
196
+ "head.bias": "final_layer.1.bias",
197
+ }
198
+ # Upstream calls the encoder ``transformer_encoder`` and the FFN linears
199
+ # ``linear1``/``linear2`` (here ``FeedForwardBlock`` children 0 and 3).
200
+ for key in self.encoder.state_dict():
201
+ upstream = key.replace("ffn.0.", "ffn.linear1.")
202
+ upstream = upstream.replace("ffn.3.", "ffn.linear2.")
203
+ self.mapping[f"transformer_encoder.{upstream}"] = f"encoder.{key}"
204
+
205
+ def reset_head(self, n_outputs: int) -> None:
206
+ self._set_n_outputs(n_outputs)
207
+ old = self.final_layer[1]
208
+ head = nn.Linear(old.in_features, n_outputs).to(old.weight)
209
+ nn.init.kaiming_uniform_(head.weight)
210
+ nn.init.zeros_(head.bias)
211
+ self.final_layer[1] = head
212
+
213
+ def forward(self, x: torch.Tensor, return_features: bool = False):
214
+ """Classify a batch of signals.
215
+
216
+ Parameters
217
+ ----------
218
+ x : torch.Tensor
219
+ Input of shape ``(batch, n_chans, n_times)``.
220
+ return_features : bool
221
+ If ``True``, return ``{"features": pooled, "cls_token": cls}``:
222
+ ``cls`` holds the class token of every channel, shape
223
+ ``(batch, n_chans, embed_dim)``, and ``pooled`` is its mean over
224
+ channels. A scripted model ignores this flag.
225
+
226
+ Returns
227
+ -------
228
+ torch.Tensor or dict
229
+ Logits of shape ``(batch, n_outputs)``, or the feature dict.
230
+ """
231
+ if x.shape[1] != self.n_chans or x.shape[2] != self.n_times:
232
+ raise ValueError(
233
+ f"EpiNT was built for (n_chans, n_times)=({self.n_chans}, "
234
+ f"{self.n_times}) and got ({x.shape[1]}, {x.shape[2]}); its LayerNorms "
235
+ "fix the number of patches."
236
+ )
237
+ x = x[..., : self.n_patches * self.patch_size]
238
+ z = self.instance_norm(self.merge_channels(x)) # (batch*chans, 1, times)
239
+ z = self.patch_embedding(self.patchify(z)) # (batch*chans, patches, dim)
240
+ cls = self.cls_token.expand(z.shape[0], 1, -1).to(z)
241
+ z = torch.cat([cls, z], dim=1) # (batch*chans, tokens, dim)
242
+ cos, sin = self.rope_cos.to(z), self.rope_sin.to(z)
243
+ for layer in self.encoder:
244
+ z = layer(z, cos, sin)
245
+ cls_token = self.split_channels(z[:, 0]) # (batch, chans, dim)
246
+ pooled = self.pool(cls_token)
247
+ if not torch.jit.is_scripting() and return_features:
248
+ return {"features": pooled, "cls_token": cls_token}
249
+ return self.final_layer(pooled)
250
+
251
+
252
+ class _RoPEAttention(nn.Module):
253
+ """Multi-head self-attention with RoPE on queries and keys."""
254
+
255
+ def __init__(self, embed_dim: int, n_heads: int):
256
+ super().__init__()
257
+ self.w_q = nn.Linear(embed_dim, embed_dim)
258
+ self.w_k = nn.Linear(embed_dim, embed_dim)
259
+ self.w_v = nn.Linear(embed_dim, embed_dim)
260
+ self.w_concat = nn.Linear(embed_dim, embed_dim)
261
+ self.split_heads = Rearrange(
262
+ "batch tokens (heads dim) -> batch heads tokens dim", heads=n_heads
263
+ )
264
+ self.merge_heads = Rearrange(
265
+ "batch heads tokens dim -> batch tokens (heads dim)"
266
+ )
267
+
268
+ def forward(
269
+ self, x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
270
+ ) -> torch.Tensor:
271
+ q = self.split_heads(self.w_q(x))
272
+ k = self.split_heads(self.w_k(x))
273
+ v = self.split_heads(self.w_v(x))
274
+ q = q * cos + rotate_pairs(q) * sin
275
+ k = k * cos + rotate_pairs(k) * sin
276
+ out = F.scaled_dot_product_attention(q, k, v)
277
+ return self.w_concat(self.merge_heads(out))
278
+
279
+
280
+ class _EpiNTLayer(nn.Module):
281
+ """Post-norm Transformer layer; the LayerNorms span tokens and features."""
282
+
283
+ def __init__(
284
+ self,
285
+ n_tokens: int,
286
+ embed_dim: int,
287
+ n_heads: int,
288
+ ffn_dim: int,
289
+ drop_prob: float,
290
+ activation: type[nn.Module],
291
+ ):
292
+ super().__init__()
293
+ self.attention = _RoPEAttention(embed_dim, n_heads)
294
+ self.norm1 = nn.LayerNorm([n_tokens, embed_dim])
295
+ self.dropout1 = nn.Dropout(drop_prob)
296
+ self.ffn = FeedForwardBlock(
297
+ emb_size=embed_dim,
298
+ expansion=1,
299
+ drop_p=drop_prob,
300
+ activation=activation,
301
+ hidden_features=ffn_dim,
302
+ output_drop_p=drop_prob,
303
+ )
304
+ self.norm2 = nn.LayerNorm([n_tokens, embed_dim])
305
+ self.dropout2 = nn.Dropout(drop_prob)
306
+
307
+ def forward(
308
+ self, x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
309
+ ) -> torch.Tensor:
310
+ x = self.norm1(x + self.dropout1(self.attention(x, cos, sin)))
311
+ return self.norm2(x + self.dropout2(self.ffn(x)))
@@ -23,6 +23,7 @@ import torch.nn.functional as F
23
23
 
24
24
  from braindecode.functional import rotate_pairs, spectral_input
25
25
  from braindecode.models.base import EEGModuleMixin
26
+ from braindecode.models.signal_jepa import _GroupNorm
26
27
  from braindecode.models.util import extract_channel_locations_from_chs_info
27
28
  from braindecode.modules.blocks import PatchTokenizer
28
29
  from braindecode.modules.layers import DropPath
@@ -884,7 +885,7 @@ class _PatchEmbedNetwork(nn.Module):
884
885
  stride=(1, self.kernel_size // 2),
885
886
  padding=(0, self.kernel_size // 2 - 1),
886
887
  ),
887
- nn.GroupNorm(self.groups, self.out_channels),
888
+ _GroupNorm(self.groups, self.out_channels),
888
889
  nn.GELU(),
889
890
  nn.Conv2d(
890
891
  in_channels=self.out_channels,
@@ -893,7 +894,7 @@ class _PatchEmbedNetwork(nn.Module):
893
894
  stride=(1, 1),
894
895
  padding=(0, 1),
895
896
  ),
896
- nn.GroupNorm(self.groups, self.out_channels),
897
+ _GroupNorm(self.groups, self.out_channels),
897
898
  nn.GELU(),
898
899
  nn.Conv2d(
899
900
  in_channels=self.out_channels,
@@ -902,7 +903,7 @@ class _PatchEmbedNetwork(nn.Module):
902
903
  stride=(1, 1),
903
904
  padding=(0, 1),
904
905
  ),
905
- nn.GroupNorm(self.groups, self.out_channels),
906
+ _GroupNorm(self.groups, self.out_channels),
906
907
  nn.GELU(),
907
908
  )
908
909
 
@@ -316,7 +316,10 @@ class _PatchEmbedding(nn.Module):
316
316
  x = self.pool(x)
317
317
  x = self.dropout(x)
318
318
  x = self.projection(x)
319
- return x.flatten(start_dim=2).transpose(1, 2)
319
+ # Contiguous tokens: each residual add takes its first operand's
320
+ # layout, so a transposed view here would keep every add and layer
321
+ # norm of the encoder reading strided memory.
322
+ return x.flatten(start_dim=2).transpose(1, 2).contiguous()
320
323
 
321
324
 
322
325
  class _ResidualAdd(nn.Module):
@@ -5,6 +5,7 @@ from typing import Tuple, Type, Union
5
5
 
6
6
  import torch
7
7
  import torch.nn as nn
8
+ import torch.nn.functional as F
8
9
  from einops.layers.torch import Rearrange
9
10
 
10
11
  from braindecode.models.base import EEGModuleMixin, _too_short_input_error
@@ -286,6 +287,120 @@ class _TSConv(nn.Sequential):
286
287
  nn.Dropout(drop_prob),
287
288
  )
288
289
 
290
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
291
+ if not torch.is_grad_enabled() and x.device.type == "hpu":
292
+ # On HPU without autograd the grouped convolution's forward is the
293
+ # faster one.
294
+ x = self[2](self[1](self[0](x)))
295
+ elif x.numel() * self[0].weight.shape[0] >= 2**26:
296
+ # ponytail: size heuristic. The spatial-first form adds ~50 small ops
297
+ # per branch; they pay off once the (batch, filters, chans, times)
298
+ # tensor it avoids reaches 2**26 elements (Gaudi lazy-mode crossover;
299
+ # eager mode and CPU gain from smaller inputs already).
300
+ x = self._spatial_first(x)
301
+ else:
302
+ # The depthwise (n_channels, 1) convolution as a matmul over the
303
+ # channel axis: same values, a faster backward than the grouped
304
+ # convolution's.
305
+ x = self[1](self[0](x))
306
+ b, f, c, t = x.shape
307
+ x = torch.matmul(self[2].weight.view(f, -1, c), x).view(b, -1, 1, t)
308
+ for i, layer in enumerate(self):
309
+ if i > 2:
310
+ x = layer(x)
311
+ return x
312
+
313
+ def _spatial_first(self, x: torch.Tensor) -> torch.Tensor:
314
+ """Temporal conv, batch norm and depthwise spatial conv (layers 0-2).
315
+
316
+ The three are linear per filter, so the spatial weights project the
317
+ input first and each filter's temporal kernel then runs on its
318
+ projections; in training the batch norm's statistics come from
319
+ :func:`_same_conv_moments`. The (batch, filters, chans, times) tensor
320
+ is never built.
321
+ """
322
+ b, _, c, t = x.shape
323
+ w = self[0].weight.flatten(1)
324
+ f, k = w.shape
325
+ w2 = self[2].weight.view(-1, c)
326
+ d = w2.shape[0] // f
327
+ if self[1].training or self[1].running_mean is None:
328
+ mean, var = _same_conv_moments(x.reshape(b * c, t), w)
329
+ if self[1].training and self[1].track_running_stats:
330
+ self[1].num_batches_tracked.add_(1)
331
+ momentum = self[1].momentum
332
+ if momentum is None:
333
+ factor = 1.0 / float(self[1].num_batches_tracked)
334
+ else:
335
+ factor = momentum
336
+ n = b * c * t
337
+ self[1].running_mean.mul_(1 - factor).add_(mean.detach(), alpha=factor)
338
+ self[1].running_var.mul_(1 - factor).add_(
339
+ var.detach(), alpha=factor * n / (n - 1)
340
+ )
341
+ else:
342
+ mean, var = self[1].running_mean, self[1].running_var
343
+ scale = self[1].weight * torch.rsqrt(var + self[1].eps)
344
+ shift = (self[1].bias - mean * scale).unsqueeze(1) * w2.sum(1).view(f, d)
345
+ y = F.conv1d(
346
+ torch.matmul(w2, x.reshape(b, c, t)),
347
+ w.unsqueeze(1).expand(f, d, k).reshape(f * d, 1, k),
348
+ padding="same",
349
+ groups=f * d,
350
+ )
351
+ scale = scale.unsqueeze(1).expand(f, d).reshape(-1, 1)
352
+ return torch.addcmul(shift.view(-1, 1), y, scale).unsqueeze(2)
353
+
354
+
355
+ def _same_conv_moments(
356
+ x: torch.Tensor, w: torch.Tensor
357
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
358
+ """Mean and biased variance per filter of ``conv1d(x, w, padding="same")``.
359
+
360
+ Over all rows and times, for ``x`` of shape (rows, times) and ``w`` of shape
361
+ (filters, k), without building the (rows, filters, times) output: the sum of
362
+ squares is the filter's autocorrelation against the input's (one matmul over
363
+ blocks of k samples), minus the outputs that "same" crops at both ends. The
364
+ input is centred first so the variance does not cancel against a large mean.
365
+ """
366
+ rows, t = x.shape
367
+ f, k = w.shape
368
+ n = rows * t
369
+ m = x.mean()
370
+ x = x - m
371
+ # a[l] = sum over rows and s of x[s] x[s + l], for l < k
372
+ nb = (t + k - 1) // k
373
+ blk = F.pad(x, (0, (nb + 1) * k - t)).view(rows, nb + 1, k)
374
+ nxt = torch.cat([blk[:, :nb], blk[:, 1:, : k - 1]], 2).reshape(-1, 2 * k - 1)
375
+ a = blk[:, :nb].reshape(-1, k).t() @ nxt
376
+ a = F.pad(a.reshape(-1), (0, k)).view(k, 2 * k)[:, :k].sum(0)
377
+ # sum of squares over every output of the full (uncropped) correlation
378
+ ww = F.conv1d(w.unsqueeze(0), w.unsqueeze(1), padding=k - 1, groups=f)[0]
379
+ sq = torch.mv(ww, torch.cat([a[1:].flip([0]), a]))
380
+ if k > 1:
381
+ right = k - 1 - (k - 1) // 2
382
+ pad = max(0, k - 1 - t)
383
+ ends = torch.cat(
384
+ [
385
+ F.pad(x[:, : k - 1], (k - 1, pad)),
386
+ F.pad(x[:, t - min(t, k - 1) :], (pad, k - 1)),
387
+ ]
388
+ )
389
+ ends = F.conv1d(ends.unsqueeze(1), w.unsqueeze(1)).view(2, rows, f, k - 1)
390
+ ends = ends.square().sum(1)
391
+ sq = sq - ends[0, :, :right].sum(1) - ends[1, :, right:].sum(1)
392
+ # undo the centring: conv(x + m) = conv(x) + m * v, v = conv(ones)
393
+ u = F.conv1d(x.sum(0).view(1, 1, t), w.unsqueeze(1), padding="same")[0]
394
+ v = F.conv1d(torch.ones_like(x[:1]).unsqueeze(0), w.unsqueeze(1), padding="same")[0]
395
+ mean_c = u.sum(1) / n
396
+ dv = v - v.mean(1, keepdim=True)
397
+ var = (
398
+ sq / n
399
+ - mean_c.square()
400
+ + m * (m * dv.square().mean(1) + 2 * (u * dv).sum(1) / n)
401
+ )
402
+ return mean_c + m * v.mean(1), var
403
+
289
404
 
290
405
  class _PositionalEncoding(nn.Module):
291
406
  r"""
@@ -16,6 +16,7 @@ from torch import Tensor, nn
16
16
 
17
17
  from braindecode.models.base import EEGModuleMixin
18
18
  from braindecode.models.labram import _Attention
19
+ from braindecode.models.signal_jepa import _GroupNorm
19
20
  from braindecode.modules import MLP, DropPath
20
21
 
21
22
  # Channel order used to train NeuroRVQ-EEG v1. Keep in sync with the released
@@ -281,7 +282,7 @@ class _MultiScaleTemporalConv(nn.Module):
281
282
  f"conv1_{i}",
282
283
  nn.Conv2d(1, out_chans, (1, kernel), padding=(0, kernel // 2)),
283
284
  )
284
- setattr(self, f"norm1_{i}", nn.GroupNorm(4, out_chans))
285
+ setattr(self, f"norm1_{i}", _GroupNorm(4, out_chans))
285
286
  setattr(self, f"pool1_{i}", nn.AvgPool2d((1, 2)))
286
287
  for i, kernel in enumerate(kernels[1], start=1):
287
288
  setattr(
@@ -289,7 +290,7 @@ class _MultiScaleTemporalConv(nn.Module):
289
290
  f"conv2_{i}",
290
291
  nn.Conv2d(out_chans, out_chans, (1, kernel), padding=(0, kernel // 2)),
291
292
  )
292
- setattr(self, f"norm2_{i}", nn.GroupNorm(4, out_chans))
293
+ setattr(self, f"norm2_{i}", _GroupNorm(4, out_chans))
293
294
  setattr(self, f"pool2_{i}", nn.AvgPool2d((1, 4)))
294
295
  self.gelu1 = activation()
295
296
  self.gelu2 = activation()
@@ -57,6 +57,7 @@ FBLightConvNet,Motor Imagery,Prediction,250,"n_chans, n_outputs, n_times, sfreq"
57
57
  IFNet,Motor Imagery,Prediction,250,"n_chans, n_outputs, n_times, sfreq",9860,"IFNet(n_chans=22, n_outputs=4, n_times=1000, sfreq=250)","Convolution,FilterBank",EEG
58
58
  BrainModule,Speech Decoding,Prediction,250,"n_chans, n_outputs, n_times, sfreq",6186909,"BrainModule(n_chans=64, n_outputs=29, n_times=160, sfreq=1000)","Convolution","EEG, MEG"
59
59
  Brant,Epilepsy,Prediction,250,"n_chans, n_outputs, n_times, sfreq",2902780,"Brant(n_chans=16, n_outputs=2, n_times=3000, sfreq=250)","Foundation Model,Attention/Transformer",iEEG
60
+ EpiNT,Epilepsy,"Prediction, Embedding",256,"n_chans, n_outputs, n_times",19194884,"EpiNT(n_chans=17, n_outputs=2, n_times=3072, sfreq=256)","Foundation Model,Attention/Transformer","EEG, iEEG"
60
61
  PBT,General,Prediction,250,"n_chans, n_outputs, n_times",818948,"PBT(n_chans=22, n_outputs=4, n_times=1000, sfreq=250)","Foundation Model",EEG
61
62
  PopulationTransformer,General,"Prediction, Embedding",2048,"n_chans, n_outputs, n_times",19967746,"PopulationTransformer(n_chans=8, n_outputs=2, n_times=768)","Foundation Model,Attention/Transformer",iEEG
62
63
  SSTDPN,Motor Imagery,Prediction,250,"n_chans, n_outputs, n_times",19502,"SSTDPN(n_chans=22, n_outputs=4, n_times=1000)","Convolution,Attention/Transformer",EEG
@@ -15,6 +15,47 @@ from braindecode.modules.linear_attention import LinearAttentionTransformer
15
15
  from braindecode.modules.quantization import EMACodebook
16
16
 
17
17
 
18
+ def _group_norm_mean_var(x, num_groups, weight, bias, eps):
19
+ """Group normalisation from mean and variance (no ``aten::group_norm``).
20
+
21
+ Same maths as :func:`torch.nn.functional.group_norm`, written with
22
+ reshape/mean/var/rsqrt ops so Intel Gaudi's graph compiler never sees
23
+ the native group-norm backward (it fails to compile it at large batch
24
+ sizes with "SRAM capping too big").
25
+ """
26
+ xg = x.reshape(x.shape[0], num_groups, -1)
27
+ mean = xg.mean(-1, keepdim=True)
28
+ var = xg.var(-1, unbiased=False, keepdim=True)
29
+ x = ((xg - mean) * torch.rsqrt(var + eps)).reshape(x.shape)
30
+ if weight is not None:
31
+ shape = (1, -1) + (1,) * (x.dim() - 2)
32
+ x = x * weight.reshape(shape)
33
+ if bias is not None:
34
+ shape = (1, -1) + (1,) * (x.dim() - 2)
35
+ x = x + bias.reshape(shape)
36
+ return x
37
+
38
+
39
+ class _GroupNorm(nn.GroupNorm):
40
+ """``nn.GroupNorm``; on HPU written out with mean and variance.
41
+
42
+ Gaudi's graph compiler fails on the native group-norm backward at large
43
+ batch sizes ("SRAM capping too big"), in both eager and lazy mode, so on
44
+ HPU the normalisation runs from mean/var ops, as
45
+ :class:`braindecode.models.signal_jepa._GroupNorm` does in HPU eager
46
+ mode. Elsewhere (CPU, CUDA; TorchScript) this is ``F.group_norm``.
47
+ """
48
+
49
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
50
+ if not torch.jit.is_scripting() and x.device.type == "hpu":
51
+ return _group_norm_mean_var(
52
+ x, self.num_groups, self.weight, self.bias, self.eps
53
+ )
54
+ return nn.functional.group_norm(
55
+ x, self.num_groups, self.weight, self.bias, self.eps
56
+ )
57
+
58
+
18
59
  class TFMTokenizerOutput(NamedTuple):
19
60
  """Outputs of :meth:`TFMTokenizer.tokenize`.
20
61
 
@@ -221,13 +262,13 @@ class TFMTokenizer(EEGModuleMixin, nn.Module, license="mit"):
221
262
  return nn.Sequential(
222
263
  nn.Conv1d(1, embed_dim, kernel_size=kernel_size, stride=stride),
223
264
  activation(),
224
- nn.GroupNorm(embed_dim // 4, embed_dim),
265
+ _GroupNorm(embed_dim // 4, embed_dim),
225
266
  nn.Conv1d(embed_dim, embed_dim, kernel_size=1),
226
267
  activation(),
227
- nn.GroupNorm(embed_dim // 4, embed_dim),
268
+ _GroupNorm(embed_dim // 4, embed_dim),
228
269
  nn.Conv1d(embed_dim, out_dim, kernel_size=1),
229
270
  activation(),
230
- nn.GroupNorm(embed_dim // 4, out_dim),
271
+ _GroupNorm(embed_dim // 4, out_dim),
231
272
  )
232
273
 
233
274
  freq_width = embed_dim // (2 * n_freq_groups)
@@ -685,6 +685,8 @@ models_mandatory_parameters: list[
685
685
  ("SSTDPN", ["n_chans", "n_outputs", "n_times", "sfreq"], None),
686
686
  ("BrainModule", ["n_chans", "n_outputs", "n_times", "sfreq"], None),
687
687
  ("Brant", ["n_chans", "n_outputs", "n_times", "sfreq"], None),
688
+ # 3072 samples, the pretrained length: 12 s at 256 Hz or 3 s at 1024 Hz
689
+ ("EpiNT", ["n_chans", "n_outputs", "n_times"], {"n_times": 3072, "sfreq": 256.0}),
688
690
  (
689
691
  "BENDR",
690
692
  ["n_chans", "n_outputs", "n_times"],
@@ -0,0 +1 @@
1
+ __version__ = "1.8.1.dev169440825"
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: braindecode
3
- Version: 1.8.1.dev169289523
3
+ Version: 1.8.1.dev169440825
4
4
  Summary: Deep learning software to decode EEG, ECG or MEG signals
5
5
  Author-email: Robin Tibor Schirrmeister <robintibor@gmail.com>, Bruno Aristimunha Pinto <b.aristimunha@gmail.com>, Alexandre Gramfort <agramfort@meta.com>
6
6
  Maintainer-email: Alexandre Gramfort <agramfort@meta.com>, Bruno Aristimunha Pinto <b.aristimunha@gmail.com>, Robin Tibor Schirrmeister <robintibor@gmail.com>
@@ -91,6 +91,7 @@ braindecode/models/eegsimpleconv.py
91
91
  braindecode/models/eegsym.py
92
92
  braindecode/models/eegtcnet.py
93
93
  braindecode/models/emg2qwerty.py
94
+ braindecode/models/epint.py
94
95
  braindecode/models/fbcnet.py
95
96
  braindecode/models/fblightconvnet.py
96
97
  braindecode/models/fbmsnet.py
@@ -179,6 +179,7 @@ interface for all EEG models and can derive variable names when needed.
179
179
  EEGSym
180
180
  EEGTCNet
181
181
  EMG2QwertyNet
182
+ EpiNT
182
183
  FBCNet
183
184
  FBLightConvNet
184
185
  FBMSNet