dataset-and-assets
Prepare, validate, and troubleshoot tf-faster-rcnn dataset folders, registry names, roidb caches, pretrained checkpoints, ImageNet initialization weights, and output/tensorboard artifact paths without downloads or training.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- dataset-and-assets
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Prepare, validate, and troubleshoot tf-faster-rcnn dataset folders, registry names, roidb caches, pretrained checkpoints, ImageNet initialization weights, and output/tensorboard artifact paths without downloads or training.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/data-layouts.md7.9 KB
- 参考文档references/model-artifacts.md6.8 KB
- 参考文档references/troubleshooting.md4.5 KB
- 脚本scripts/validate_layout.py11.0 KB
03 SKILL.md 正文
dataset-and-assets
Use this sub-skill when a future Researcher needs to place or validate VOC/COCO data, understand dataset registry names, reason about roidb and data/cache, place pretrained Faster R-CNN checkpoints or ImageNet initialization weights, or locate generated output and TensorBoard artifacts for tf-faster-rcnn.
Do not use this sub-skill for dependency installation, Cython/CUDA build failures, TensorFlow compatibility, or NMS extension import errors; route those to installation-and-configuration. Do not include or construct training/test/demo commands here; route command construction and runtime choices to training-and-evaluation or inference-and-demo.
Operating inputs
- A checkout root, referenced below as
<repo-root>. - An intended dataset family: PASCAL VOC, COCO, the bundled
data/demoimages, pretrained Faster R-CNN checkpoint assets, and/or ImageNet initialization weights. - Whether the user wants a read-only validation of paths or instructions for arranging existing local/downloaded assets. Never download from this sub-skill without explicit user approval; the bundled validator performs no network access.
Exact dataset registry keys
lib/datasets/factory.py registers 23 keys. Use these exact strings for --imdb values and for roidb/cache names:
- VOC 2007:
voc_2007_train,voc_2007_val,voc_2007_trainval,voc_2007_test - VOC 2012:
voc_2012_train,voc_2012_val,voc_2012_trainval,voc_2012_test - VOC 2007 difficult-inclusive variants:
voc_2007_train_diff,voc_2007_val_diff,voc_2007_trainval_diff,voc_2007_test_diff - VOC 2012 difficult-inclusive variants:
voc_2012_train_diff,voc_2012_val_diff,voc_2012_trainval_diff,voc_2012_test_diff - COCO 2014:
coco_2014_train,coco_2014_val,coco_2014_minival,coco_2014_valminusminival,coco_2014_trainval - COCO 2015:
coco_2015_test,coco_2015_test-dev
voc_2007_trainval+voc_2012_trainval and coco_2014_train+coco_2014_valminusminival are not registry keys accepted directly by get_imdb; they are combined training-set strings split by tools/trainval_net.py. Route command use of those combined strings to training-and-evaluation.
Asset workflow
-
Establish
<repo-root>and treat<repo-root>/dataascfg.DATA_DIR. -
Pick the registry key(s) or source-supported combined training string needed by the downstream task.
-
Validate the intended layout with the bundled script:
python <skill-dir>/scripts/validate_layout.py --repo-root <repo-root> --check voc python <skill-dir>/scripts/validate_layout.py --repo-root <repo-root> --check coco python <skill-dir>/scripts/validate_layout.py --repo-root <repo-root> --check demo-model python <skill-dir>/scripts/validate_layout.py --repo-root <repo-root> --check imagenetOmit
--checkto run all layout checks. The script reports deterministic JSON and exits nonzero when required expected paths for the selected check are absent. -
If data annotations or image-set files changed, remove stale
data/cache/<imdb-name>_gt_roidb.pklfiles before the next dataset constructor run. -
For pretrained checkpoint placement and output/TensorBoard path reasoning, use
references/model-artifacts.md.
Reference map
references/data-layouts.md: VOC/COCO folder contracts, exact registry keys, image-name patterns,roidbdict fields, cache semantics, and safe validation examples.references/model-artifacts.md: pretrained Faster R-CNN checkpoint layout,data/imagenet_weightsnaming, snapshot sidecars, demo-model symlink layout, and output/TensorBoard path formulas.references/troubleshooting.md: common missing-path, stale-cache, split-file, checkpoint-sidecar, and route-to-other-sub-skill decisions.scripts/validate_layout.py: safe stdlib-only path/name validator; no imports from the legacy repo, no downloads, no dataset parsing, no TensorFlow execution.
Guardrails
- Do not run
data/scripts/fetch_faster_rcnn_models.shautomatically. It downloads a large archive and depends on an old external server. - Do not assert full COCO/VOC data correctness from path checks alone; this sub-skill validates expected paths and names, not annotation content.
- Do not treat CUDA/NVCC/TensorFlow errors as dataset problems. Route install/build/backend failures to
installation-and-configuration. - Do not generate training/test commands in this sub-skill. Route commands, iterations, config overrides, snapshots, resume behavior, and benchmark evaluation to
training-and-evaluation.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).