Skip to content

Effect chart

Per-seed paired deltas of the default and rewire arms in the seven-generation matrix: backbone on the x axis, the same-seed paired delta on the y axis. Above the zero line the grafted block won that seed.

Fig. 1Paired deltas across seven generationsdots for seeds, bar for the mean, line for the range
Data: results/summary.md

The bar is the mean, drawn only with three seeds or more (the v10n default arm has five); the vertical line is the range and the dots are the seeds. Switch between mAP50 and mAP50-95 above the chart, hover a cell for its details, and save a PNG from the corner.

How to read it

  • Read the range, not only the bar. Three seeds often straddle zero even when the mean is positive; the judgment lines said up front that three seeds support no significance test, and even 3/3 gives a sign-test p of 0.125.
  • The trend of the means is the main finding: the default wiring is positive on SPPF-ended backbones (v5n, v8n, v9t), sits on zero once the backbone ends in attention (v10n, 11n), and turns negative with area attention and the E2E head (12n, 26n).
  • An arm with fewer than three seeds gets no mean, only its dots.
  • The two arms have the same parameter count (3,327,330 on YOLOv8n); only the wiring differs.

Area buckets

Fig. 2Split by object sizepaired deltas in COCO area buckets, bars for the mean of three seeds
Data: results/buckets.md

The buckets show no consistent size trade-off: with the default wiring on v8n large objects get worse on all three seeds (APl βˆ’0.0104), and rewire turns that positive; on 26n it is small objects that lose (APs βˆ’0.0045, 0/3).

Alignment arms

Fig. 3Alignment armsupstream's settings inside the block, four blocks and balance terms, each against the same-seed baseline
Data: results/summary.md

These arms compare block configurations on one backbone, which Fig. 1's generation axis cannot carry. Output normalisation and dense training are both 3/3 on v5n; four blocks are 0/3 on both v10n and v5n. Round-by-round verdicts are on Experiments and Judgment lines.

Recompute

uv run python scripts/report.py   # paired tables
uv run python scripts/charts.py   # figures and data

scripts/charts.py computes the numbers from results/*.json into docs/javascripts/data.js, the same source as Results and Judgment lines. The SVGs under docs/assets/ are static versions of the same data, used by the README and the wiki.