📦 WP-022 — feat(models): dual detection head, reg_max=1¶
Research log · Phase 3 — Assignment and losses
🔗 Depends on WP-021 · Needed by WP-023, WP-041
WP-022 — the dead predictor¶
The mechanism and its fix are A30. What belongs here is the failure signature, because it is the most misleading one this project has produced.
With zero-initialized class-conv biases every sigmoid opens at 0.5, the TAL-normalized summed BCE starts near 1.8M nats, the first MuSGD step at lr 0.01 blows the weights out (box predictions reached ±6000 px), TAL then assigns no positives, and the network saturates to an exact-zero loss with zero gradient. The progress bar shows train/loss 0.000 — which reads as a perfectly converged model rather than a dead one. Reproduced identically on CPU and MPS from real COCO batches.
Post-fix verification was a 30-step real-COCO check showing monotone descent 14k → 1.5k with box predictions bounded within image scale — i.e. the fix was confirmed on the same signal that had lied, not on "the tests pass".