same 23→72.83 endpoints, two different percentage-gain claims in the posters. ugh recompute from the raw numbers; keep scoring definition + denominator nearby. percentage points and percent gain are not the same thing Engineering-plan targets...
poster recipe: 3 conv, 2 ReLU, flatten, 2 dense; lr 0.001 / 128 epochs / L2 missing widths, batch size, L2 coeff + exact loss path. save the checkpoint/log next to the diagram if found The poster supplies more concrete model details than the...
513 checks → 115, nice need runtime + input size too. fewer visited voxels doesn't automatically mean the same factor of speedup The tiny complexity sheet records exhaustive = 513 and tracing = 115 iterations. The ratio is 513/115 ≈ 4.46 ;...
script → artifact map, otherwise i'm going to lose track of what produced what FAISS loss sketch detaches to numpy... where are the gradients supposed to go? keep embedding training separate until that is answered Stage Script named in...
RecobundlesX config pile: lengths / votes / seed / processes / transforms older requirements fighting newer deps, then headers lost in conversion. keep one coherent set of inputs for the next run Recorded configuration: length filtering minL=20 ,...
PRE.mif vs IN.mif again... right next to 'header transformations ... do not match' check template + correspondence + directions/index files. renaming it won't fix the space it's in pipeline context: step 19 writes fiber-density correspondence to...
SIFT COMPLETE — 20 million → 2 million keep FOD + mask + both tractograms together. SIFT filtering and SIFT2 weights are different outputs, don't mix filenames files: the command operates in the template directory. Keep FOD.mif ,...
reading pile: bottlenecks / false positives / FBC / phantom validation question to come back to — what ambiguity survives even when the synthetic example looks convincing? Saved readings worth keeping next to the implementation:...
fixel index first, voxel coords second — multiple directions can live at the same xyz need one attribute order, stop hardcoding the reshape. tiny fake grid with 2 crossing bundles might be easier to debug than an entire brain representation:...
pretty tractogram ≠ correct connections, annoying but useful reminder keep validity definition with the score. especially when changing clustering / endpoint voting problem: a streamline's visual plausibility is not enough to establish that it...
pipeline versions: 20M vs 25M tracks, 0.06 vs 0.1 cutoff, TractSeg vs RecoBundlesX keep separate configs. subject/template transforms need to follow every output The Random Notes workflow starts with response estimation using Dhollander, a...
regression NOT classification — density / angle / position for matching in→out pieces what if several candidates want the same exit? also try tracing the bottleneck backward, should the pairings agree? The scratch note explicitly says...
