Repository navigation
I-ML-01: 官方 COCO80 与 PP-OCRv6 CPU 推理及可重建证据 - #12
Merged
Merged
Conversation
Deliver batches 31-35 with the supplied official ONNX artifacts: preserve all 80 COCO classes, run PP-OCRv6_small on CPU, gate complete image batches with real quality metrics, and rebuild OCR crops from shared recipes. Record artifact/runtime hashes and synthetic replay evidence; keep model weights outside Git. No CUDA or production activation is introduced. Validation: 54 CPU adapter/process/pixel tests passed without skips; real three-image inference reproduced all nine evidence crops including three rotated recognition regions. Contract generation, design validation and Ruff passed; fixture HTTP tests ran in the business test environment. Rollback: revert the model adapter/schema/configuration changes and use the prior development fixture profile. Do not relabel old v4/four-class manifests or evaluation evidence, and do not activate unapproved bundles.
10 of 13 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
本次交付第31–35批：直接使用已提供的官方D-FINE COCO80 ONNX和PP-OCRv6_small det/rec ONNX，在CPU进行真实质量检查、检测和文字识别。OCR实际识别像素来自共享裁剪配方，可跨进程重建核SHA；补拍或技术失败无部分成功结果。
任务：I-ML-01；关联AI-05/06/07/09/10与UP-02的本地工程子范围，现场评测/完整业务接线/生产批准保持未关闭。
PR顺序
本PR以
wsq/i-02-remediation-reports为基线，使第27–30批业务改动保持在独立PR中。先审查并Squash合并整改/报告PR #11，再将本PR基线切到main并同步分支、重新执行检查；不要将两个任务的差异混在同一个审查范围。变更文件
apps/ai_inference/adapters/{dfine,ocrv6,quality}.py：80类、固定制品SHA、张量/provider/数值/几何门禁、真实质量算法与OCR。apps/ai_inference/{cpu,ocr_cpu,pipeline_cpu}.py及CLI：1–3图、有界spawn/kill/join、全批质量门禁、检测/文字各100上限，唯一CPUExecutionProvider。packages/image_evidence/perspective.py：共享perspective-rgb-v1、RGB PNG、源RGB/PNG/识别像素摘要和独立旋转，OCR报告升级v2。tools/design/{dfine_reference,extra_specs,release_specs,...}.py及生成契约：COCO80、ocr_backend=onnxruntime、ocrv6smallcpu runtime profile；不手改生成物。tools/mlsmoke和docs/08-delivery/31–35及合成证据同步。验证
docs/08-delivery/35-crop-cpu-evidence.json。--check、设计发布43/D-FINE112/readiness148及文档链接、Ruff check/format、pip check与diff检查通过。风险、审批与回滚
COCO四类到80类、Paddle OCR到ONNX OCR和旧runtime profile均为未部署设计的兼容性变化；本地OCR裁剪改为黑色边界，adapter/本地报告升级v2，旧验收证据保留。按
CONTRIBUTING.md，模型/契约变更须两位协作者共同确认，独立批准并通过CI后Squash合并。当前仅CPU、不安装CUDA；HTTP仍mock，真实常驻服务、受控analysis对象读取/版本核验、OCR文字的RPC引用闭包、Worker证据写入、瓶子/化学/日期事实与现场准确率未完成。合成图成绩不能用于安全判断或生产批准。COCO没有label/shelf/cabinet，不伪造这些检测类别。
回滚适配器/生成源与契约/依赖配置，恢复先前development fixture profile；保持生产拒绝未批准bundle。禁止将旧四类/v4制品或报告改名复用。模型、凭据、真实图片、虚拟环境和忽略目录不提交。