Skip to content

perf: widen powerset window hop to 2.0s (~2.5× faster segmentation) - #83

Merged
ekhodzitsky merged 1 commit into
masterfrom
pr/7-powerset-hop-2
Jul 31, 2026
Merged

perf: widen powerset window hop to 2.0s (~2.5× faster segmentation)#83
ekhodzitsky merged 1 commit into
masterfrom
pr/7-powerset-hop-2

Conversation

@ekhodzitsky

@ekhodzitsky ekhodzitsky commented Jul 30, 2026

Copy link
Copy Markdown
Owner

Что

Дефолтный hop powerset-окон 1.0 с → 2.0 с (перекрытие 90% → 80%): вдвое меньше 10-секундных окон на файл при той же агрегации. Затронуты дефолт PowersetConfig.hop_secs, обе записи в src/models/manifest.toml, пересчитан DER-бейзлайн ami_test, добавлены бенч-артефакты в benchmarks/results/powerset-hop2-2026-07-30/.

Валидация (pipeline v2 + VBx, CPU provider, полные корпуса)

корпус DER был → стал segmentation RTFx
AMI-test (16 файлов) 17.66 → 15.71 551 → 207 с (×2.7) 31 → 68
VoxConverse-dev (216 файлов) 7.90 → 7.70 976 → 453 с (×2.2) 38 → 56
AMI EN2002a (single) 24.76 → 24.27

DER не деградировал ни на одном корпусе — на AMI даже −1.95 п.п. (меньше избыточного перекрытия → чище агрегированные постериоры).

Проверки

  • 459 lib-тестов, data-gated DER-гейты (der_baseline / der_v2_baseline / cli_der_regression) — все зелёные
  • clippy чист; v2_e2e_smoke и long-form floor в пределах бейзлайнов, не менялись
  • Полный voxconverse-test (232) с hop 2.0 не перемерян — vox-dev улучшился, риск низкий; можно отдельным прогоном

Halves the number of overlapping 10-s segmentation windows per file,
cutting segmentation time by 2.2-2.7x with no DER regression (pipeline
v2 + VBx, CPU provider):

- AMI-test 16 meetings: DER 17.66 -> 15.71 (collar 0.25, micro),
  segmentation 551 -> 207 s, RTFx 31 -> 68
- VoxConverse-dev 216 files: DER 7.90 -> 7.70, segmentation
  976 -> 453 s, RTFx 38 -> 56
- AMI EN2002a single meeting: DER 24.76 -> 24.27

Updates the PowersetConfig default and the shipped manifest geometry
(both powerset_fp32 and powerset_int8 entries), re-measures the
ami_test DER baseline, and records the fresh benchmark artifacts.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@ekhodzitsky
ekhodzitsky force-pushed the pr/7-powerset-hop-2 branch from 2bb6f12 to 524919d Compare July 31, 2026 07:47
@ekhodzitsky
ekhodzitsky merged commit 3df6132 into master Jul 31, 2026
47 checks passed
@ekhodzitsky
ekhodzitsky deleted the pr/7-powerset-hop-2 branch July 31, 2026 07:55
ekhodzitsky added a commit that referenced this pull request Jul 31, 2026
perf: widen powerset window hop to 2.0s (~2.5× faster segmentation)
ekhodzitsky added a commit that referenced this pull request Jul 31, 2026
perf: widen powerset window hop to 2.0s (~2.5× faster segmentation)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant