Skip to content

REVIEW.md

Latest commit

 

History

History
81 lines (74 loc) · 3.22 KB

REVIEW.md

File metadata and controls

81 lines (74 loc) · 3.22 KB

LibreYOLO Review Guide

Use this file as context for agents performing PR reviews. This is not general implementation guidance for agents writing code. Keep review comments specific, evidence-based, and scoped to the PR under review.

Core axioms

  • Humans own issues, PRs, reviews, merges.
  • One PR solves one problem.
  • Shared changes need blast-radius justification.
  • Keep PRs as small as possible.
  • Read /docs before reviewing contracts.
  • Metadata is the model-loading source of truth.
  • Filenames are not model identity.
  • State-dict sniffing is legacy compatibility.
  • Foreign weights require converter scripts.
  • Official checkpoints use flat v1 metadata.
  • Checkpoints store state dicts, not objects.
  • Metadata changes update docs and helpers.
  • License compatibility is non-negotiable.
  • One factory routes model families.
  • Family constants define names, sizes, tasks.
  • Task is a first-class axis.
  • Detect is the suffixless default task.
  • Task resolution is explicit, metadata, suffix, default.
  • Cross-family checkpoint loads fail.
  • Cross-task checkpoint loads fail.
  • Not every family supports every task.
  • YOLO9 and RF-DETR anchor coverage.
  • Public APIs follow Ultralytics-style conventions.
  • Explicit user kwargs beat defaults.
  • CLI defaults are family-derived.
  • Config dataclasses define training truth.
  • Trainers orchestrate; families own recipes.
  • RF-DETR ignores generic YOLO augmentation knobs.
  • RF-DETR learning rate is absolute.
  • DDP batch means global batch.
  • Per-rank loaders divide global batch.
  • Python multi-GPU training auto-spawns DDP.
  • Torchrun owns rank and device environment.
  • DDP loss scales by world size.
  • Rank zero owns side effects.
  • Autobatch returns rank-divisible global batches.
  • Unit tests prove CPU-safe API behavior.
  • Install smoke proves clean environment importability.
  • GPU nightly proves real-model behavior.
  • Nightly-selected skips are failures.
  • Original-canvas coordinates are canonical.
  • YOLO labels feed COCO metrics.
  • Preprocessing is family-local.
  • Validation is shared and task-aware.
  • Results stay flat and API-compatible.
  • Backends must behave like models.
  • Exported runtimes round-trip metadata.
  • DETR outputs use top-k, not NMS.
  • Segmentation metrics are mask-first.
  • Pose validation uses COCO OKS semantics.

Review focus

  • Flag unrelated changes bundled into bugfix PRs.
  • Flag shared-code changes from model-specific bugs.
  • Flag metadata behavior that bypasses /docs.
  • Flag filename heuristics replacing metadata contracts.
  • Flag API behavior that silently accepts ignored options.
  • Flag changes that weaken normal single-GPU training.
  • Flag DDP fixes that regress non-DDP paths.
  • Flag CI marker changes that hide tests.
  • Flag heavyweight tests in the fast unit suite.
  • Flag GPL, AGPL, LGPL, proprietary, or unknown derivations.
  • Flag PR descriptions that omit meaningful behavior changes.

Contract references

  • docs/checkpoint_schema.md: checkpoint metadata and loading rules.
  • docs/nomenclature.md: model names, tasks, suffixes, resolution order.
  • docs/testing.md: test tiers and validation expectations.
  • docs/adr/: architecture decisions and design contracts.
  • CONTRIBUTING.md: contribution and metadata-change policy.
  • AGENTS.md: agent and PR policy.