wwu-mmll / confound_corrected_cpm

86%

main: 85%

Build Type

push

github

Committed by NilsWinter

Commit Message

Brain-plot selector, connected-component selection, GPU benchmark + VRAM fix

Features
- Brain & Edges report: default to significantly stable edges (NBS/TFCE p<0.05)
  and add a self-contained button group to switch the matrix/hub/chord views
  between Significant / Top 5% / Top 10% by stability (masked_signed_stability_
  matrix + inline JS/CSS). Glass brain renders the default subset.
- Edge selection: `UnivariateEdgeSelection(connected_components=True|int)` keeps
  only edges in a connected component with >= N edges (per network), dropping
  isolated edges to improve stability. Applied per fold and per permutation.
- scripts/benchmark_cpu_gpu.py: times edge selection + model fit on CPU vs CUDA
  across sizes/permutations. Shows CPU≈GPU is expected at perms=1 (overhead
  dominates the tiny matmuls); the GPU only wins with large permutation batches
  and/or big connectomes.

Fixes
- GPU OOM in edge-stability aggregation. ResultsManager preallocated the per-fold
  edge masks [Features, 2, Folds, Runs] on the compute device and densified them
  to node×node arrays there, needing tens of GB of VRAM for large parcellations ×
  many folds × many permutations. Edge bookkeeping now lives on the CPU as a
  running fold-sum (all stability needs); node×node arrays are built on the CPU;
  per-fold edges.npy is kept for the real run only (the permutation pass keeps
  just the fold-averaged stability_edges.npy for the null). Verified on GPU:
  100 nodes × 10 folds × 200 perms now peaks at ~69 MB VRAM for this step.

Docs: methods.md (presence/connected-component filters), interpreting_results.md
(brain selector), installation.md (limited-VRAM guidance), CHANGELOG, RELEASE_PLAN.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Coverage Stats

110 of 118 new or added lines in 4 files covered. (93.22%)

43 existing lines in 5 files now uncovered.

2374 of 2750 relevant lines covered (86.33%)

0.86 hits per line