braindecode 1.8.1.dev178155531__tar.gz → 1.8.1.dev178224482__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 (208) hide show
  1. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/NOTICE.txt +4 -0
  2. {braindecode-1.8.1.dev178155531/braindecode.egg-info → braindecode-1.8.1.dev178224482}/PKG-INFO +1 -1
  3. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/augmentation/__init__.py +2 -0
  4. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/augmentation/functional.py +6 -2
  5. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/augmentation/transforms.py +185 -0
  6. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/base.py +4 -1
  7. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/tuh.py +12 -9
  8. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/functional/__init__.py +4 -0
  9. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/functional/functions.py +114 -10
  10. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/__init__.py +2 -0
  11. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/attn_sleep.py +5 -11
  12. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/axon.py +15 -8
  13. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/base.py +10 -1
  14. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/bendr.py +28 -5
  15. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/biot.py +25 -15
  16. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/brainbert.py +2 -8
  17. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/brainmodule.py +6 -11
  18. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/brainomni.py +79 -44
  19. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/brant.py +5 -4
  20. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/cbramod.py +22 -20
  21. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/codebrain.py +87 -67
  22. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/csbrain.py +31 -15
  23. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/dance.py +1 -1
  24. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegdino.py +10 -10
  25. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eeginception_mi.py +44 -17
  26. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegpt.py +47 -74
  27. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegsym.py +17 -1
  28. braindecode-1.8.1.dev178224482/braindecode/models/guetschel2026.py +775 -0
  29. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/labram.py +59 -43
  30. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/luna.py +34 -30
  31. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/mapa.py +116 -73
  32. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/medformer.py +24 -19
  33. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/meta_neuromotor.py +45 -48
  34. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/mirepnet.py +7 -2
  35. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/msvtnet.py +7 -6
  36. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/mvpformer.py +172 -191
  37. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/neurorvq.py +42 -21
  38. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/neurorvq_tokenizer.py +88 -47
  39. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/patchedtransformer.py +0 -1
  40. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/reve.py +38 -52
  41. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/signal_jepa.py +35 -15
  42. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sleepfm.py +56 -70
  43. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sstdpn.py +13 -10
  44. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/steegformer.py +6 -4
  45. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/summary.csv +1 -0
  46. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/tcformer.py +25 -48
  47. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/tfm_tokenizer.py +8 -15
  48. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/util.py +143 -34
  49. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/attention.py +4 -4
  50. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/dance_modules.py +34 -26
  51. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/filter.py +3 -14
  52. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/layers.py +12 -21
  53. braindecode-1.8.1.dev178224482/braindecode/modules/linear_attention.py +128 -0
  54. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/parametrization.py +29 -0
  55. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/quantization.py +49 -39
  56. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/preprocessing/preprocess.py +9 -0
  57. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/preprocessing/windowers.py +5 -4
  58. braindecode-1.8.1.dev178224482/braindecode/version.py +1 -0
  59. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482/braindecode.egg-info}/PKG-INFO +1 -1
  60. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode.egg-info/SOURCES.txt +2 -0
  61. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/api.rst +7 -0
  62. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/user_guide/channel_strategies.rst +1 -1
  63. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/whats_new.rst +122 -0
  64. braindecode-1.8.1.dev178155531/braindecode/version.py +0 -1
  65. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/LICENSE.txt +0 -0
  66. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/MANIFEST.in +0 -0
  67. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/README.rst +0 -0
  68. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/__init__.py +0 -0
  69. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/augmentation/base.py +0 -0
  70. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/classifier.py +0 -0
  71. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/__init__.py +0 -0
  72. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/_notebook_viewer.py +0 -0
  73. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bbci.py +0 -0
  74. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bcicomp.py +0 -0
  75. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/__init__.py +0 -0
  76. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/datasets.py +0 -0
  77. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/format.py +0 -0
  78. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/hub.py +0 -0
  79. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/hub_format.py +0 -0
  80. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/hub_io.py +0 -0
  81. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/hub_validation.py +0 -0
  82. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/bids/iterable.py +0 -0
  83. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/chb_mit.py +0 -0
  84. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/collate.py +0 -0
  85. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/mne.py +0 -0
  86. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/moabb.py +0 -0
  87. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/nmt.py +0 -0
  88. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/registry.py +0 -0
  89. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/siena.py +0 -0
  90. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/sleep_physio_challe_18.py +0 -0
  91. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/sleep_physionet.py +0 -0
  92. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/utils.py +0 -0
  93. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datasets/xy.py +0 -0
  94. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datautil/__init__.py +0 -0
  95. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datautil/channel_utils.py +0 -0
  96. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datautil/hub_formats.py +0 -0
  97. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datautil/serialization.py +0 -0
  98. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/datautil/util.py +0 -0
  99. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/eegneuralnet.py +0 -0
  100. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/functional/_real_dft.py +0 -0
  101. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/functional/initialization.py +0 -0
  102. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/atcnet.py +0 -0
  103. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/attentionbasenet.py +0 -0
  104. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/barista.py +0 -0
  105. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/config.py +0 -0
  106. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/contrawr.py +0 -0
  107. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/ctnet.py +0 -0
  108. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/deep4.py +0 -0
  109. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/deepsleepnet.py +0 -0
  110. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/dgcnn.py +0 -0
  111. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/diver1.py +0 -0
  112. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eeg_clip.py +0 -0
  113. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegconformer.py +0 -0
  114. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eeginception_erp.py +0 -0
  115. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegitnet.py +0 -0
  116. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegminer.py +0 -0
  117. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegnet.py +0 -0
  118. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegnex.py +0 -0
  119. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegsimpleconv.py +0 -0
  120. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/eegtcnet.py +0 -0
  121. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/emg2qwerty.py +0 -0
  122. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/fbcnet.py +0 -0
  123. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/fblightconvnet.py +0 -0
  124. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/fbmsnet.py +0 -0
  125. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/hybrid.py +0 -0
  126. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/ifnet.py +0 -0
  127. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/mscformer.py +0 -0
  128. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/neuropose.py +0 -0
  129. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/popt.py +0 -0
  130. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sccnet.py +0 -0
  131. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/seizure_transformer.py +0 -0
  132. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sensingdynamics.py +0 -0
  133. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/shallow_fbcsp.py +0 -0
  134. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sinc_shallow.py +0 -0
  135. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sleep_stager_blanco_2020.py +0 -0
  136. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sleep_stager_chambon_2018.py +0 -0
  137. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/sparcnet.py +0 -0
  138. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/syncnet.py +0 -0
  139. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/tcn.py +0 -0
  140. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/tidnet.py +0 -0
  141. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/tmsanet.py +0 -0
  142. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/tsinception.py +0 -0
  143. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/usleep.py +0 -0
  144. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/vemg2pose.py +0 -0
  145. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/models/zuna.py +0 -0
  146. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/__init__.py +0 -0
  147. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/activation.py +0 -0
  148. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/blocks.py +0 -0
  149. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/channels.py +0 -0
  150. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/convolution.py +0 -0
  151. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/linear.py +0 -0
  152. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/popt_modules.py +0 -0
  153. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/stats.py +0 -0
  154. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/util.py +0 -0
  155. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/modules/wrapper.py +0 -0
  156. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/preprocessing/__init__.py +0 -0
  157. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/preprocessing/eegprep_preprocess.py +0 -0
  158. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/preprocessing/mne_preprocess.py +0 -0
  159. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/preprocessing/util.py +0 -0
  160. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/regressor.py +0 -0
  161. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/samplers/__init__.py +0 -0
  162. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/samplers/base.py +0 -0
  163. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/samplers/ssl.py +0 -0
  164. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/training/__init__.py +0 -0
  165. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/training/callbacks.py +0 -0
  166. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/training/losses.py +0 -0
  167. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/training/scoring.py +0 -0
  168. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/util.py +0 -0
  169. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/__init__.py +0 -0
  170. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/activations.py +0 -0
  171. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/attribution.py +0 -0
  172. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/confusion_matrices.py +0 -0
  173. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/frequency.py +0 -0
  174. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/metrics.py +0 -0
  175. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/sanity.py +0 -0
  176. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode/visualization/topology.py +0 -0
  177. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode.egg-info/dependency_links.txt +0 -0
  178. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode.egg-info/requires.txt +0 -0
  179. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/braindecode.egg-info/top_level.txt +0 -0
  180. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/Makefile +0 -0
  181. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/_templates/autosummary/class.rst +0 -0
  182. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/_templates/autosummary/class_in_subdir.rst +0 -0
  183. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/_templates/autosummary/function.rst +0 -0
  184. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/_templates/autosummary/function_in_subdir.rst +0 -0
  185. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/agent_docs.rst +0 -0
  186. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/cite.rst +0 -0
  187. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/conf.py +0 -0
  188. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/help.rst +0 -0
  189. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/index.rst +0 -0
  190. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/install/install.rst +0 -0
  191. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/install/install_pip.rst +0 -0
  192. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/install/install_source.rst +0 -0
  193. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/attention.rst +0 -0
  194. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/channel.rst +0 -0
  195. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/convolution.rst +0 -0
  196. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/filterbank.rst +0 -0
  197. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/gnn.rst +0 -0
  198. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/interpretable.rst +0 -0
  199. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/lbm.rst +0 -0
  200. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/recurrent.rst +0 -0
  201. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/categorization/spd.rst +0 -0
  202. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/models.rst +0 -0
  203. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/models_categorization.rst +0 -0
  204. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/models_table.rst +0 -0
  205. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/models/models_visualization.rst +0 -0
  206. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/docs/sg_execution_times.rst +0 -0
  207. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/pyproject.toml +0 -0
  208. {braindecode-1.8.1.dev178155531 → braindecode-1.8.1.dev178224482}/setup.cfg +0 -0
@@ -49,14 +49,18 @@ The following components are licensed under the MIT License:
49
49
  - `braindecode/modules/popt_modules.py`
50
50
  - `braindecode/models/brainomni.py`
51
51
  - `braindecode/modules/quantization.py`
52
+ - `braindecode/modules/linear_attention.py`
52
53
  - `braindecode/models/seizure_transformer.py`
53
54
  - `docs/_static/model/seizure_transformer_arch.png`
55
+ - `braindecode/models/guetschel2026.py`
54
56
  - `braindecode/models/tmsanet.py`
55
57
 
56
58
  The `brainomni.py` file is a derivative of `OpenTSLab/BrainOmni` (MIT), whose
57
59
  SEANet codec derives from Meta's `facebookresearch/encodec` (MIT). The
58
60
  `quantization.py` module holds the residual vector quantization derived from
59
61
  `lucidrains/vector-quantize-pytorch` (MIT) and the BrainOmni port.
62
+ The `linear_attention.py` module is the part of
63
+ `lucidrains/linear-attention-transformer` (MIT) that BIOT and TFMTokenizer run.
60
64
 
61
65
  ### Apache-2.0 Licensed Files
62
66
 
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: braindecode
3
- Version: 1.8.1.dev178155531
3
+ Version: 1.8.1.dev178224482
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>
@@ -23,6 +23,7 @@ from .transforms import (
23
23
  SignFlip,
24
24
  SmoothTimeMask,
25
25
  TimeReverse,
26
+ TrivialAugment,
26
27
  )
27
28
 
28
29
  __all__ = [
@@ -30,6 +31,7 @@ __all__ = [
30
31
  "IdentityTransform",
31
32
  "Compose",
32
33
  "AugmentedDataLoader",
34
+ "TrivialAugment",
33
35
  "TimeReverse",
34
36
  "SignFlip",
35
37
  "FTSurrogate",
@@ -3,6 +3,7 @@
3
3
  # Gustavo Rodrigues <gustavenrique01@gmail.com>
4
4
  # Bruna Lopes <brunajaflopes@gmail.com>
5
5
  # Sarthak Tayal <sarthaktayal2@gmail.com>
6
+ # Anton Soloviev <anton@praviel.com>
6
7
  #
7
8
  # License: BSD (3-clause)
8
9
 
@@ -931,7 +932,7 @@ def _rotate_signals(
931
932
  transformed_X = X.clone()
932
933
  sensors_positions = list(sensors_positions_matrix)
933
934
  for s, rot_sensors_matrix in enumerate(rot_sensors_matrices):
934
- rot_sensors_positions = list(rot_sensors_matrix.T)
935
+ rot_sensors_positions = list(rot_sensors_matrix)
935
936
  for time in range(X.shape[-1]):
936
937
  interpolator_t = Rbf(*sensors_positions, X[s, :, time])
937
938
  transformed_X[s, :, time] = torch.from_numpy(
@@ -1093,7 +1094,8 @@ def segmentation_reconstruction(
1093
1094
  y : torch.Tensor
1094
1095
  EEG labels for the example or batch.
1095
1096
  n_segments : int
1096
- Number of segments to use in the batch.
1097
+ Number of segments to use in the batch. Any remaining samples are
1098
+ included in the last segment.
1097
1099
  data_classes : list[tuple[int, torch.Tensor]]
1098
1100
  List of tuples. Each tuple contains the class index and the corresponding EEG data.
1099
1101
  rand_indices : array-like
@@ -1133,6 +1135,8 @@ def segmentation_reconstruction(
1133
1135
  for idx_segment in range(n_segments):
1134
1136
  start = idx_segment * segment_size
1135
1137
  end = (idx_segment + 1) * segment_size
1138
+ if idx_segment == n_segments - 1:
1139
+ end = window_size
1136
1140
 
1137
1141
  # Perform the data augmentation
1138
1142
  X_aug[np.arange(n_trials), :, start:end] = X_class[
@@ -1457,3 +1457,188 @@ class BandRotation(Transform):
1457
1457
  "circular_jitter": self.circular_jitter,
1458
1458
  "random_state": self.rng,
1459
1459
  }
1460
+
1461
+
1462
+ class TrivialAugment(Transform):
1463
+ """Apply one randomly sampled label-preserving augmentation per example.
1464
+
1465
+ For every example, independently samples one transform from the pool
1466
+ below and one of ``num_bins`` strengths evenly spaced between that
1467
+ transform's bounds, as in TrivialAugment [1]_. Examples that draw the
1468
+ same transform and strength are transformed in one call; each pooled
1469
+ transform still draws its own randomness per example.
1470
+
1471
+ Default pool (no assumption on the signal's unit, sampling frequency or
1472
+ montage):
1473
+
1474
+ * :class:`TimeReverse` — no strength parameter.
1475
+ * :class:`SignFlip` — no strength parameter.
1476
+ * :class:`FTSurrogate` — phase noise magnitude, 0.2 to 1.
1477
+ * :class:`ChannelsDropout` — channel drop probability, 0.05 to 0.4.
1478
+ * :class:`ChannelsShuffle` — shuffle probability, 0.1 to 0.75.
1479
+ * :class:`SmoothTimeMask` — mask length as a fraction of the window
1480
+ length, 0.05 to 0.25.
1481
+ * :class:`GaussianNoise` — noise standard deviation as a fraction of the
1482
+ sub-batch standard deviation (volts or microvolts alike), 0.02 to 0.3.
1483
+ * :class:`AmplitudeScale` — scale interval ``(1/s, s)``, ``s`` from 1.25
1484
+ to 2.5.
1485
+
1486
+ ``sfreq`` adds :class:`BandstopFilter` (bandwidth 0.5 to 4 Hz, capped
1487
+ 2.5 Hz below Nyquist) and :class:`FrequencyShift` (maximum shift 0.5 to
1488
+ 2 Hz); ``sensors_positions_matrix`` adds :class:`SensorsRotation` around
1489
+ the z axis (5 to 25 degrees).
1490
+
1491
+ Parameters
1492
+ ----------
1493
+ probability : float, optional
1494
+ Probability of augmenting each example. Defaults to 1.0.
1495
+ sfreq : float | None, optional
1496
+ Sampling frequency of the signals, in Hz. Defaults to None.
1497
+ sensors_positions_matrix : array-like | None, optional
1498
+ Sensor positions of shape ``(3, n_channels)`` (see
1499
+ :class:`SensorsRotation`). Defaults to None.
1500
+ num_bins : int, optional
1501
+ Number of strength levels per transform. Defaults to 10.
1502
+ random_state : int | numpy.random.RandomState, optional
1503
+ Seed shared by the per-example sampler and every pooled transform.
1504
+ Defaults to None.
1505
+
1506
+ Attributes
1507
+ ----------
1508
+ op_names : list of str
1509
+ Names of the transforms in the pool, in sampling order.
1510
+
1511
+ References
1512
+ ----------
1513
+ .. [1] Müller, S. G., & Hutter, F. (2021). TrivialAugment:
1514
+ Tuning-free Yet State-of-the-Art Data Augmentation. Proceedings of
1515
+ the IEEE/CVF International Conference on Computer Vision (ICCV),
1516
+ pp. 774-782.
1517
+ """
1518
+
1519
+ def __init__(
1520
+ self,
1521
+ probability=1.0,
1522
+ sfreq=None,
1523
+ sensors_positions_matrix=None,
1524
+ num_bins=10,
1525
+ random_state=None,
1526
+ ):
1527
+ super().__init__(probability=probability, random_state=random_state)
1528
+ if not isinstance(num_bins, (int, np.integer)) or num_bins < 1:
1529
+ raise ValueError(f"num_bins must be a positive integer, got {num_bins}")
1530
+ self.num_bins = int(num_bins)
1531
+ rng = self.rng
1532
+ # (name, strength bounds or None, make(strength) or None: built per call)
1533
+ specs = [
1534
+ ("TimeReverse", None, lambda s: TimeReverse(1.0, random_state=rng)),
1535
+ ("SignFlip", None, lambda s: SignFlip(1.0, random_state=rng)),
1536
+ (
1537
+ "FTSurrogate",
1538
+ (0.2, 1.0),
1539
+ lambda s: FTSurrogate(1.0, phase_noise_magnitude=s, random_state=rng),
1540
+ ),
1541
+ (
1542
+ "ChannelsDropout",
1543
+ (0.05, 0.4),
1544
+ lambda s: ChannelsDropout(1.0, p_drop=s, random_state=rng),
1545
+ ),
1546
+ (
1547
+ "ChannelsShuffle",
1548
+ (0.1, 0.75),
1549
+ lambda s: ChannelsShuffle(1.0, p_shuffle=s, random_state=rng),
1550
+ ),
1551
+ ("SmoothTimeMask", (0.05, 0.25), None),
1552
+ ("GaussianNoise", (0.02, 0.3), None),
1553
+ (
1554
+ "AmplitudeScale",
1555
+ (1.25, 2.5),
1556
+ lambda s: AmplitudeScale(1.0, interval=(1.0 / s, s), random_state=rng),
1557
+ ),
1558
+ ]
1559
+ if sfreq is not None:
1560
+ nyquist = sfreq / 2.0
1561
+ # BandstopFilter needs bandwidth < nyquist - 2; keep a margin
1562
+ max_bandwidth = min(4.0, nyquist - 2.5)
1563
+ if max_bandwidth <= 0.5:
1564
+ raise ValueError(
1565
+ f"sfreq={sfreq} is too small to sample valid band-stop "
1566
+ "bandwidths (needs at least 6.0 Hz)."
1567
+ )
1568
+ specs += [
1569
+ (
1570
+ "BandstopFilter",
1571
+ (0.5, max_bandwidth),
1572
+ lambda s: BandstopFilter(
1573
+ 1.0,
1574
+ sfreq=sfreq,
1575
+ bandwidth=s,
1576
+ max_freq=nyquist,
1577
+ random_state=rng,
1578
+ ),
1579
+ ),
1580
+ (
1581
+ "FrequencyShift",
1582
+ (0.5, 2.0),
1583
+ lambda s: FrequencyShift(
1584
+ 1.0, sfreq=sfreq, max_delta_freq=s, random_state=rng
1585
+ ),
1586
+ ),
1587
+ ]
1588
+ if sensors_positions_matrix is not None:
1589
+ specs.append(
1590
+ (
1591
+ "SensorsRotation",
1592
+ (5.0, 25.0),
1593
+ lambda s: SensorsRotation(
1594
+ 1.0,
1595
+ sensors_positions_matrix=sensors_positions_matrix,
1596
+ max_degrees=s,
1597
+ random_state=rng,
1598
+ ),
1599
+ )
1600
+ )
1601
+
1602
+ self._pool = []
1603
+ for name, bounds, make in specs:
1604
+ if bounds is None:
1605
+ strengths = [None]
1606
+ else:
1607
+ strengths = np.linspace(*bounds, self.num_bins).tolist()
1608
+ variants = None if make is None else [make(s) for s in strengths]
1609
+ self._pool.append((name, strengths, variants))
1610
+ # pre-built transforms as submodules (pickle, ``.to()``)
1611
+ self._static_transforms = torch.nn.ModuleList(
1612
+ v for _, _, variants in self._pool if variants for v in variants
1613
+ )
1614
+ self.op_names = [name for name, _, _ in self._pool]
1615
+
1616
+ def operation(self, X, y):
1617
+ """Apply one sampled transform and strength to each example of X."""
1618
+ n_strengths = np.array([len(strengths) for _, strengths, _ in self._pool])
1619
+ op_ids = self.rng.randint(0, len(self._pool), size=X.shape[0])
1620
+ bin_ids = self.rng.randint(0, self.num_bins, size=X.shape[0])
1621
+ # magnitude-free transforms collapse all examples onto bin 0
1622
+ bin_ids[n_strengths[op_ids] == 1] = 0
1623
+ keys = op_ids * self.num_bins + bin_ids
1624
+
1625
+ out_X, out_y = X.clone(), y.clone()
1626
+ for key in np.unique(keys):
1627
+ group = torch.from_numpy(np.flatnonzero(keys == key)).to(X.device)
1628
+ op_i, bin_j = divmod(int(key), self.num_bins)
1629
+ name, strengths, variants = self._pool[op_i]
1630
+ X_sub = X[group]
1631
+ if variants is not None:
1632
+ transform = variants[bin_j]
1633
+ elif name == "SmoothTimeMask":
1634
+ n_mask = max(1, int(round(strengths[bin_j] * X.shape[-1])))
1635
+ transform = SmoothTimeMask(
1636
+ 1.0, mask_len_samples=n_mask, random_state=self.rng
1637
+ )
1638
+ else: # GaussianNoise, std relative to the sub-batch (unit-free)
1639
+ std = strengths[bin_j] * X_sub.std().item()
1640
+ transform = GaussianNoise(1.0, std=std, random_state=self.rng)
1641
+ tr_X, tr_y = transform(X_sub, y[group])
1642
+ out_X[group] = tr_X.to(out_X.dtype)
1643
+ out_y[group] = tr_y
1644
+ return out_X, out_y
@@ -1742,7 +1742,10 @@ class BaseConcatDataset(ConcatDataset, HubDatasetMixin, Generic[T]):
1742
1742
  # Create path if it doesn't exist
1743
1743
  os.makedirs(path, exist_ok=True)
1744
1744
 
1745
- path_contents = os.listdir(path)
1745
+ # With an offset the caller fills one directory piece by piece (preprocess
1746
+ # saves each recording at its index), so the other entries are expected:
1747
+ # skip the scan that would stat them all on every call and list them in a warning.
1748
+ path_contents = os.listdir(path) if offset == 0 else []
1746
1749
  n_sub_dirs = len(
1747
1750
  [e for e in path_contents if os.path.isdir(os.path.join(path, e))]
1748
1751
  )
@@ -11,10 +11,12 @@ TUH Abnormal EEG Corpus.
11
11
  from __future__ import annotations
12
12
 
13
13
  import glob
14
+ import json
14
15
  import os
15
16
  import re
16
17
  import warnings
17
18
  from datetime import datetime, timezone
19
+ from functools import cache
18
20
  from pathlib import Path
19
21
  from typing import Iterable, Literal
20
22
  from unittest import mock
@@ -183,11 +185,8 @@ class TUH(BaseConcatDataset):
183
185
  }
184
186
  raw.rename_channels(mapping_strip)
185
187
 
186
- montage1005 = mne.channels.make_standard_montage(
187
- resolve_montage_name("standard_1005")
188
- )
189
188
  mapping_eeg_names = {
190
- c.upper(): c for c in montage1005.ch_names if c.upper() in raw.ch_names
189
+ c.upper(): c for c in _standard_1005().ch_names if c.upper() in raw.ch_names
191
190
  }
192
191
 
193
192
  # Set channels whose type could not be inferred (defaulted to "eeg") to "misc":
@@ -210,10 +209,7 @@ class TUH(BaseConcatDataset):
210
209
 
211
210
  @staticmethod
212
211
  def _set_montage(raw):
213
- montage = mne.channels.make_standard_montage(
214
- resolve_montage_name("standard_1005")
215
- )
216
- raw.set_montage(montage, on_missing="ignore")
212
+ raw.set_montage(_standard_1005(), on_missing="ignore")
217
213
 
218
214
  @staticmethod
219
215
  def _create_dataset(
@@ -269,6 +265,12 @@ class TUH(BaseConcatDataset):
269
265
  return base_dataset
270
266
 
271
267
 
268
+ @cache
269
+ def _standard_1005():
270
+ # Built once per process: set_montage copies the montage before using it.
271
+ return mne.channels.make_standard_montage(resolve_montage_name("standard_1005"))
272
+
273
+
272
274
  def _create_description(file_paths, version, ds_name):
273
275
  descriptions = [
274
276
  _parse_description_from_file_path(f, version, ds_name) for f in file_paths
@@ -289,7 +291,8 @@ def _read_date(file_path):
289
291
  date_path = file_path.replace(".edf", "_date.txt")
290
292
  # if date file exists, read it
291
293
  if os.path.exists(date_path):
292
- description = pd.read_json(date_path, typ="series").to_dict()
294
+ with open(date_path) as f:
295
+ description = json.load(f)
293
296
  # otherwise read edf file, extract date and store to file
294
297
  else:
295
298
  raw = mne.io.read_raw_edf(file_path, preload=False, verbose="error")
@@ -3,11 +3,13 @@ from .functions import (
3
3
  daubechies_filters,
4
4
  drop_path,
5
5
  dwt_max_level,
6
+ fft_conv1d,
6
7
  hilbert_freq,
7
8
  identity,
8
9
  iou_1d,
9
10
  pairwise_iou_1d,
10
11
  plv_time,
12
+ prefer_fft_conv,
11
13
  rotate_pairs,
12
14
  safe_log,
13
15
  sinusoidal_positional_encoding,
@@ -22,11 +24,13 @@ __all__ = [
22
24
  "daubechies_filters",
23
25
  "drop_path",
24
26
  "dwt_max_level",
27
+ "fft_conv1d",
25
28
  "hilbert_freq",
26
29
  "identity",
27
30
  "iou_1d",
28
31
  "pairwise_iou_1d",
29
32
  "plv_time",
33
+ "prefer_fft_conv",
30
34
  "rotate_pairs",
31
35
  "safe_log",
32
36
  "sinusoidal_positional_encoding",
@@ -7,6 +7,7 @@
7
7
  from __future__ import annotations
8
8
 
9
9
  import math
10
+ from typing import Optional
10
11
 
11
12
  import numpy as np
12
13
  import torch
@@ -54,6 +55,95 @@ def spectral_input(x: torch.Tensor) -> torch.Tensor:
54
55
  return x.to(torch.promote_types(x.dtype, torch.float32))
55
56
 
56
57
 
58
+ def fft_conv1d(
59
+ x: torch.Tensor, weight: torch.Tensor, bias: Optional[torch.Tensor] = None
60
+ ) -> torch.Tensor:
61
+ """``F.conv1d(x, weight, bias, padding="same")`` computed with FFTs.
62
+
63
+ Equal to the direct convolution up to float rounding: the FFT is long
64
+ enough for a linear (not circular) convolution. Its cost barely grows with
65
+ the kernel size, so it is faster than the direct one for long kernels on
66
+ CPU (see :func:`prefer_fft_conv`). float16/bfloat16 inputs are computed in
67
+ float32 (:func:`spectral_input`) and cast back.
68
+
69
+ Parameters
70
+ ----------
71
+ x : torch.Tensor
72
+ Input of shape ``(batch, in_channels, n_times)``.
73
+ weight : torch.Tensor
74
+ Kernels of shape ``(out_channels, in_channels, kernel_size)``.
75
+ bias : torch.Tensor | None
76
+ Bias of shape ``(out_channels,)``.
77
+
78
+ Returns
79
+ -------
80
+ torch.Tensor
81
+ Output of shape ``(batch, out_channels, n_times)``, dtype and device of
82
+ ``x``. For an even ``kernel_size`` it is aligned like torch's
83
+ ``padding="same"`` (the extra zero padded on the right).
84
+ """
85
+ n_times, kernel_size = x.shape[-1], weight.shape[-1]
86
+ # >= n_times + kernel_size - 1 (no circular wrap-around), even and
87
+ # 5-smooth: the fast cases of a real FFT
88
+ n_fft = 2 * _next_fast_len((n_times + kernel_size) // 2)
89
+ xs = spectral_input(x)
90
+ # conv1d is a cross-correlation: a convolution with the flipped kernel
91
+ x_f = torch.fft.rfft(xs, n=n_fft).permute(2, 0, 1) # (freq, batch, in)
92
+ w_f = torch.fft.rfft(weight.to(xs).flip(-1), n=n_fft).permute(2, 1, 0)
93
+ # Per frequency (batch, in) @ (in, out) in complex numbers, as one real
94
+ # matmul (a complex one is slower on CPU): [Re x; Im x] @ [Re w, Im w]
95
+ # (columns interleaved) holds the four real products.
96
+ p = torch.cat([x_f.real, x_f.imag], 1) @ torch.view_as_real(w_f).flatten(2)
97
+ p = p.unflatten(1, [2, x.shape[0]]).unflatten(3, [-1, 2]) # x re/im, w re/im
98
+ y_f = torch.complex(
99
+ p[:, 0, :, :, 0] - p[:, 1, :, :, 1], p[:, 0, :, :, 1] + p[:, 1, :, :, 0]
100
+ )
101
+ y = torch.fft.irfft(y_f.permute(1, 2, 0), n=n_fft)
102
+ # "same" output t is sample t + kernel_size // 2 of the full convolution
103
+ y = y[..., kernel_size // 2 : kernel_size // 2 + n_times]
104
+ if bias is not None:
105
+ y = y + bias.to(y).unsqueeze(-1)
106
+ return y.to(x)
107
+
108
+
109
+ def prefer_fft_conv(x: torch.Tensor, kernel_size: int) -> bool:
110
+ """Whether :func:`fft_conv1d` is expected to beat a direct convolution.
111
+
112
+ Rule measured for :class:`~braindecode.models.EEGInceptionMI` (48 to 240
113
+ channels, 128-500 Hz, 2 CPU threads, oneDNN on and off): on CPU the FFT
114
+ wins for float16/bfloat16 inputs (no fast half-precision CPU convolution)
115
+ and from ``batch_size * kernel_size >= 600`` if also
116
+ ``in_channels * kernel_size >= 120``: with one or two input channels a
117
+ float32 direct convolution stays cheaper up to k = 85 / 31 (13 long
118
+ temporal convolutions of the model zoo, 22 channels x batch 32). GPUs and
119
+ HPUs (no complex dtype) keep the direct convolution.
120
+
121
+ Parameters
122
+ ----------
123
+ x : torch.Tensor
124
+ Input of the convolution, ``(batch, in_channels, n_times)``.
125
+ kernel_size : int
126
+ Kernel length in samples.
127
+ """
128
+ return x.device.type == "cpu" and (
129
+ x.element_size() == 2
130
+ or (x.shape[0] * kernel_size >= 600 and x.shape[1] * kernel_size >= 120)
131
+ )
132
+
133
+
134
+ def _next_fast_len(n: int) -> int:
135
+ """Smallest ``2^a 3^b 5^c >= n`` (scipy's ``next_fast_len(n, real=True)``)."""
136
+ n -= 1
137
+ m = 0
138
+ while m != 1:
139
+ n += 1
140
+ m = n
141
+ for p in [2, 3, 5]:
142
+ while m % p == 0:
143
+ m = m // p
144
+ return n
145
+
146
+
57
147
  def drop_path(
58
148
  x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
59
149
  ):
@@ -197,6 +287,7 @@ def hilbert_freq(x: torch.Tensor, forward_fourier: bool = True) -> torch.Tensor:
197
287
  return _real_dft.hilbert_freq_real(x, forward_fourier).to(x.dtype)
198
288
 
199
289
  input_dtype = x.dtype
290
+ cast_back = x.is_floating_point()
200
291
  # ``view_as_complex`` does not accept bfloat16 real/imaginary pairs. The
201
292
  # HPU path can produce bfloat16 Fourier coefficients under autocast, so do
202
293
  # the complex-valued part in float32 and restore the real-valued contract
@@ -224,9 +315,8 @@ def hilbert_freq(x: torch.Tensor, forward_fourier: bool = True) -> torch.Tensor:
224
315
  x = torch.fft.ifft(x, norm=None, dim=-1) # returns complex signal
225
316
  x = torch.view_as_real(x)
226
317
 
227
- if input_dtype == torch.bfloat16:
228
- x = x.to(input_dtype)
229
- return x
318
+ # Integer input stays float (casting back would truncate the analytic signal).
319
+ return x.to(input_dtype) if cast_back else x
230
320
 
231
321
 
232
322
  def plv_time(
@@ -281,18 +371,24 @@ def plv_time(
281
371
  )
282
372
  # Normalize the analytic signal to obtain unit vectors (phasors).
283
373
  unit_phasor = analytic_signal / amplitude.unsqueeze(-1)
374
+ # The real/imaginary slices are strided (step 2): matmul copied each of
375
+ # them for every product, per batch matrix. Copy them once.
376
+ real = unit_phasor[..., 0].contiguous()
377
+ imag = unit_phasor[..., 1].contiguous()
378
+ real_t = real.transpose(-2, -1).contiguous()
379
+ imag_t = imag.transpose(-2, -1).contiguous()
284
380
 
285
381
  # Compute the real part of the outer product between phasors of
286
382
  # different channels.
287
- real_real = torch.matmul(unit_phasor[..., 0], unit_phasor[..., 0].transpose(-2, -1))
383
+ real_real = torch.matmul(real, real_t)
288
384
 
289
385
  # Compute the imaginary part of the outer product between phasors of
290
386
  # different channels.
291
- imag_imag = torch.matmul(unit_phasor[..., 1], unit_phasor[..., 1].transpose(-2, -1))
387
+ imag_imag = torch.matmul(imag, imag_t)
292
388
 
293
389
  # Compute the cross-terms for the real and imaginary parts.
294
- real_imag = torch.matmul(unit_phasor[..., 0], unit_phasor[..., 1].transpose(-2, -1))
295
- imag_real = torch.matmul(unit_phasor[..., 1], unit_phasor[..., 0].transpose(-2, -1))
390
+ real_imag = torch.matmul(real, imag_t)
391
+ imag_real = torch.matmul(imag, real_t)
296
392
 
297
393
  # Combine the real and imaginary parts to form the complex correlation.
298
394
  correlation_real = real_real + imag_imag
@@ -412,7 +508,14 @@ def dwt_max_level(n_times: int, filter_len: int) -> int:
412
508
  """
413
509
  if n_times < filter_len:
414
510
  return 0
415
- return max(0, int(math.floor(math.log2(n_times / (filter_len - 1)))))
511
+ # floor(log2(n_times / (filter_len - 1))) in integers (TorchScript has no
512
+ # math.log2): the largest level with (filter_len - 1) * 2**level <= n_times.
513
+ ratio = n_times // (filter_len - 1)
514
+ level = 0
515
+ while ratio >= 2:
516
+ ratio = ratio // 2
517
+ level += 1
518
+ return level
416
519
 
417
520
 
418
521
  def wavelet_decomposition(
@@ -456,8 +559,9 @@ def wavelet_decomposition(
456
559
  )
457
560
  approx, detail = F.conv1d(padded, weight, stride=2).unbind(dim=1)
458
561
  details.append(detail)
459
- out = torch.cat([approx, *reversed(details)], dim=-1)
460
- return out.reshape(*leading, -1)
562
+ details.reverse() # [cD_n, ..., cD_1]
563
+ out = torch.cat([approx] + details, dim=-1)
564
+ return out.reshape(list(leading) + [-1])
461
565
 
462
566
 
463
567
  def sinusoidal_positional_encoding(
@@ -39,6 +39,7 @@ from .emg2qwerty import EMG2QwertyNet
39
39
  from .fbcnet import FBCNet
40
40
  from .fblightconvnet import FBLightConvNet
41
41
  from .fbmsnet import FBMSNet
42
+ from .guetschel2026 import Guetschel2026
42
43
  from .hybrid import HybridNet
43
44
  from .ifnet import IFNet
44
45
  from .labram import Labram
@@ -139,6 +140,7 @@ __all__ = [
139
140
  "FBCNet",
140
141
  "FBLightConvNet",
141
142
  "FBMSNet",
143
+ "Guetschel2026",
142
144
  "MetaNeuromotorHand",
143
145
  "HybridNet",
144
146
  "IFNet",
@@ -407,16 +407,12 @@ class _MRCNN(nn.Module):
407
407
  ##########################################################################################
408
408
 
409
409
 
410
- def _attention(
411
- query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
412
- ) -> tuple[torch.Tensor, torch.Tensor]:
413
- """Implementation of Scaled dot product attention."""
410
+ def _attention(query: torch.Tensor, key: torch.Tensor) -> torch.Tensor:
411
+ """Scaled dot product attention weights."""
414
412
  # d_k - dimension of the query and key vectors
415
413
  d_k = query.size(-1)
416
414
  scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_k)
417
- p_attn = F.softmax(scores, dim=-1) # attention weights
418
- output = torch.matmul(p_attn, value) # (B, h, T, d_k)
419
- return output, p_attn
415
+ return F.softmax(scores, dim=-1) # attention weights
420
416
 
421
417
 
422
418
  class _MultiHeadedAttention(nn.Module):
@@ -464,10 +460,8 @@ class _MultiHeadedAttention(nn.Module):
464
460
  .transpose(1, 2)
465
461
  )
466
462
 
467
- x_raw, attn_weights = _attention(query, key, value)
468
- # apply dropout to the *weights*
469
- attn = self.dropout(attn_weights)
470
- # recompute the weighted sum with dropped weights
463
+ # apply dropout to the *weights*, then the weighted sum
464
+ attn = self.dropout(_attention(query, key))
471
465
  x = torch.matmul(attn, value)
472
466
 
473
467
  # merge heads and project
@@ -363,8 +363,11 @@ class _AXONEncoder(nn.Module):
363
363
  position = spatial.unsqueeze(1) + temporal.unsqueeze(0) # (C, P, D)
364
364
  tokens = tokens + position.unsqueeze(0).to(tokens.dtype)
365
365
 
366
- index = torch.arange(n_patches, device=tokens.device)
367
- band = (index.unsqueeze(0) - index.unsqueeze(1)).abs() <= self.temporal_window
366
+ # Up to temporal_window + 1 patches the band covers every pair: no mask.
367
+ band: Optional[torch.Tensor] = None
368
+ if n_patches > self.temporal_window + 1:
369
+ index = torch.arange(n_patches, device=tokens.device)
370
+ band = (index[None] - index[:, None]).abs() <= self.temporal_window
368
371
 
369
372
  for block in self.blocks:
370
373
  tokens = block(tokens, band)
@@ -464,19 +467,23 @@ class _AxisAttention(nn.Module):
464
467
  )
465
468
  return attended.transpose(1, 2).reshape(n_seq, length, -1)
466
469
 
467
- def forward(self, x: torch.Tensor, band: torch.Tensor) -> torch.Tensor:
470
+ def forward(self, x: torch.Tensor, band: Optional[torch.Tensor]) -> torch.Tensor:
468
471
  batch, n_chans, n_patches, dim = x.shape
469
472
  axis_weights = self.axis_gate(x) # (B, C, P, 2)
470
473
  scale_weights = self.scale_gate(x) # (B, C, P, 2)
471
474
 
472
- # Temporal path: one sequence of P tokens per electrode.
475
+ # Temporal path: one sequence of P tokens per electrode. Without a band
476
+ # the restricted branch is the full one; attend once unless the two
477
+ # branches draw their own attention dropout.
473
478
  per_channel = self.temporal_qkv(x.reshape(batch * n_chans, n_patches, dim))
474
479
  restricted = self.temporal_out(self._attend(per_channel, band)).reshape(
475
480
  batch, n_chans, n_patches, dim
476
481
  )
477
- full = self.temporal_out(self._attend(per_channel, None)).reshape(
478
- batch, n_chans, n_patches, dim
479
- )
482
+ full = restricted
483
+ if band is not None or (self.training and self.att_drop_prob > 0):
484
+ full = self.temporal_out(self._attend(per_channel, None)).reshape(
485
+ batch, n_chans, n_patches, dim
486
+ )
480
487
  temporal = scale_weights[..., 0:1] * restricted + scale_weights[..., 1:2] * full
481
488
 
482
489
  # Spatial path: one sequence of C tokens per time step.
@@ -511,6 +518,6 @@ class _AXONBlock(nn.Module):
511
518
  self.norm_ffn = nn.RMSNorm(embed_dim, eps=1e-8)
512
519
  self.ffn = _GatedFeedForward(embed_dim, ffn_expansion)
513
520
 
514
- def forward(self, x: torch.Tensor, band: torch.Tensor) -> torch.Tensor:
521
+ def forward(self, x: torch.Tensor, band: Optional[torch.Tensor]) -> torch.Tensor:
515
522
  x = x + self.attn(self.norm_attn(x), band)
516
523
  return x + self.ffn(self.norm_ffn(x))