EgoHOS: rename shared-v1.0 to shared-v1.1 and replace doc/CONTRACT.md (candidate PR)
Candidate PR: rename EgoHOS shared-v1.0 to shared-v1.1 and replacedoc/CONTRACT.md with the v1.1 contract. Please review before merging.
EgoHOS is the one dataset in this migration with nothing to merge. Its registry declares
exactly two components, rgb and segment, and both are media; the only .pt files in the
tree are the 212 media validity files (rgb_valid.pt / mask_valid.pt, 360,114 B in
total), which v1.1 keeps exactly as they are. There is no per-frame numeric tensor file
anywhere under EgoHOS/shared-v1.0, so v1.1's merge rules bind nothing here, no reader
needs adapting, and there is no local migration or acceptance run to make. What remains is
the version change itself: the directory rename and the contract.
What this PR changes
EgoHOS/shared-v1.1is created fromEgoHOS/shared-v1.0by server-side copy. All
1,570 files areCommitOperationCopyoperations sourced atmain, so every byte is
preserved and none of the 3.25 GiB is re-uploaded or re-encoded. This is a rename, not a
rebuild -- the plan is built entirely from the published tree's own index.doc/CONTRACT.mdbecomes the v1.1 contract (15,323 B). It is built from the house
draft by the same three-edit recipe ARCTIC's and RHD's published contracts use: a dataset
banner, a dataset status paragraph, and the draft's link into the local draft directory
rewritten toV1_0_SEMANTICS.md(plus CRLF to LF and the trailing newline dropped, as in
both siblings). The build script proves the recipe rather than asserting it: it rebuilds
ARCTIC's and RHD's contracts from the same draft and requires the results to be
byte-identical to what those releases publish.doc/V1_0_SEMANTICS.md(45,148 B) is the previous content ofdoc/CONTRACT.md,
published under the name the v1.1 contract links to. It costs no upload -- it is a copy of
a file already in the tree. This is whydoc/CONTRACT.mdis the single path in this plan
copied to a different destination.doc/MIGRATION.md(6,205 B) is added: byte-identical to the copy ARCTIC and RHD
publish, and linked five times from the contract shipped here, so shipping the contract
without it would ship dangling links..gitattributes(see below) carries EgoHOS's.mkvpaths in glob form.EgoHOS/rawis untouched, and no other dataset's root or LFS tracking line is
modified -- both re-checked on the candidate branch, not assumed.
Only two files are uploaded, ~22 KB. main keeps shared-v1.0 fully available until this
PR merges.
Registered deviations -- the reviewer should decide whether they are acceptable
This PR does not claim the dataset was adapted to v1.1, and the shipped contract's own
status paragraph says so explicitly. The contract requires metainfo.schema_version to end
up at hand-warehouse/1.1 and the registry to match, and warns that changing version
strings -- or re-badging the v1.0 registry -- does not make a dataset compliant. Three
things are therefore left as they are, on purpose, and named rather than implied:
- all 106
data/*/ego/metainfo.jsonstill declareschema_version: hand-warehouse/1.0,release_id: shared-v1.0; doc/schema.jsonis still the v1.0 registry (the contract links to it twice);script/is not adapted --script/CONTRACT.mdandscript/schema.jsonare still the
v1.0 documents.
If the reviewer would rather have (1) and (2) done as well, that is a small follow-up on
this same branch (106 small JSON files plus a 39,043 B registry); it is being raised here
precisely because the honest version of "rename plus contract" is what this PR is.
File counts
Unlike ARCTIC and RHD, this dataset gets no file-count reduction -- there was nothing to
merge. The tree goes from 1,570 files to 1,572: the two documents above. The motivation
here is version consistency across the warehouse, not the Hub's 1,000,000-file limit;
EgoHOS was never near it.
.gitattributes
This repo tracks LFS with a mix of global patterns (*.pt, *.png, *.jpg, ...) and
explicit per-path lines. At main that file is 10,484,654 B against the Hub's 10 MiB
ceiling for non-LFS files -- 1,106 B of headroom. .mkv has no global pattern, so the Hub
appends one tracking line per new .mkv path: this release's 212 of them would need about
23 KB and cannot fit. EgoHOS's own 156 lines in that file are exactly its v1.0 .mkv paths
(17,498 B; no other EgoHOS suffix is tracked by a per-path line), so this PR replaces them
with two globs, exactly the device the repository already carries fordex-ycb/shared-v1.1/**/*.obj on main, for HOT3D / OakInk-v1 / OakInk-v2 / FreiHAND in
PR #4, and for ARCTIC and RHD in their candidate PRs -- leaving about 18 KB of headroom.
With the old tree removed the file keeps only the EgoHOS/shared-v1.1/**/*.mkv glob:
deliberately not the shape of stripping the dataset's lines outright, because the Hub
appended no per-path lines for this release (it went up while the glob was live) and that
line is what keeps the 212 .mkv LFS-tracked for a fresh clone. Every non-EgoHOS line is
byte-identical, re-checked on the candidate branch before merging.
The glob commit is the PR's first commit on purpose: staging any path before it would let
the Hub append its per-path lines and push the file over the ceiling -- ARCTIC's release was
refused exactly that way on its first attempt.
Verification performed on the candidate branch before this was pushed
Exact file set; per-file size and content identity (LFS sha256 or git blob id) against the
published v1.0 index; every copy's source still present and unchanged in the old tree;EgoHOS/raw tree id unchanged; the old release untouched until it was removed in one
separate commit after all of the above passed; no other dataset's root changed; no
non-EgoHOS LFS line changed; .gitattributes under the Hub's ceiling.
Pushed under yhl's authorization. Merging is left to the reviewer -- the
tool that pushed this branch has no merge option at all, deliberately.
🤖 Generated with Claude Code
复核意见的修改已经推上本候选分支(未合并)。
| 项 | 值 |
|---|---|
| head | 9a5b1f286c24c8ddd91f245ce39462df3a509963 → 671194f1c18121fe9f415f813665ab0c99c58e4a |
EgoHOS/shared-v1.1 树 |
a13a98da… → c5c077c7f7b2485637f1521d4c84d2d3e0e2fdce |
| 文件 | 1,571(1,572 − 1) |
| 改动 | 123 改 / 0 增 / 1 删 / 2,507,905 B(2.4 MiB) |
推法是最小重推:本地树与分支自己的文件索引(发布时记下的每个路径的 size 与内容身份)逐路径对比,只提交真正不同的字节——其余 1,447 个路径一个字节都没有重传。
改的 123 个 = 106 个视图的 metainfo.json + 17 个 doc/script 文件:
- 106 个
metainfo.json的字节差恰好是schema_version(→hand-warehouse/1.1)与release_id(→shared-v1.1)两行,别的不动; - 版本与工具适配:
preprocess.py的版本串与上传目标改为从schema.json读(不再硬编码),两个校验器各加check_release_versions(发布目录名 / 跨视图版本一致 /build.json的 recipe 摘要必须描述本树这些工具的字节)与check_view_coverage(build['views']必须等于树上的data/*/*/metainfo.json数、且非空——也就是「0 个视图也返回通过」那一类洞)。
删的 1 个:doc/validation/human_review.json。它是 v1.0 那次构建的人工批准记录:它记的 report_sha256 与 recipe.code_sha256 指的是 v1.0 的工具。v1.1 重新验收之后这份批准已经作废(对不上号的批准比没有更糟),而本发布按 human_review: pending 出厂——所以明确删掉,并把理由登记在规范 §3 的登记里,而不是留一份会误导人的通过记录。原文保留在树外(tmp/egohos-tools/v1_0_backup/doc/validation/),随时可比。
没动的内容:3.2 GB 媒体与其余 1,447 个路径,逐条从 Hub 重读核对,仍带已发布 v1.0 的身份——1,108 条按 LFS 的内容 sha256 比(Hub 对这类路径存的是指针,树清单给出的 git blob id 是那个指针的 id,与内容不可比),339 条普通文件按 git blob id 比。
本树的验收重跑过:106 视图 / 11,743 帧、两份注册表统一到 hand-warehouse/1.1 / shared-v1.1,清单 1,571 文件 / 3,249,544,332 B;树外回归测试 17 条全过。规范 §3 的登记在 tmp/egohos-tools/MIGRATION_REGISTRATION.json(源 v1.0 的 schema/固定提交/清单摘要、目标版本、工具摘要、逐路径改动与删除清单、checkpoint、阶段、错误、验收结果)。独立复查(另起进程、另走路子读 Hub)tmp/egohos-tools/verify_pr7_state.py → PR7_STATE_OK。
仍停在合并之前:human_full_review: pending、publication_authorized: false、hf_writes: false 一律未动。
另一件需要你定的事:Hub 提示本 PR 与 main 有一个 .gitattributes 冲突——main 在我们复核期间被别的数据集推着前移了。这不是本 release 的缺陷(发布器验的是「相对本 PR 的 base commit 只动了 EgoHOS 自己那几行」,成立;本 PR 把 EgoHOS 的 155 条逐路径 LFS 行并成了 glob),但合并前需要把两边合起来。处理办法与 RHD PR #6、ARCTIC PR #5 同一类(以当前 main 为底套用本 release 自己的那几处改动,算出文件但不推送、交你定),要不要我们也算一份,交你定。
Review fixes applied and tested; current main LFS rules reconciled without changing other datasets or raw data. Regression checks pass and content manifests refreshed. Data identity is preserved; historical semantic exceptions and pending human review remain explicitly documented in doc/REVIEW_FIXES.md. This merge does not constitute human visual acceptance.