AGENT SKILL

data-and-datasets

Operate FastReID dataset registries, built-in layouts, custom dataset classes, dataloaders, transforms, samplers, and safe data preflight checks.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
data-and-datasets
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Operate FastReID dataset registries, built-in layouts, custom dataset classes, dataloaders, transforms, samplers, and safe data preflight checks.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

FastReID data and datasets

Use this sub-skill when a FastReID task involves dataset roots, FASTREID_DATASETS, built-in dataset names, custom dataset registration, ImageDataset/CommDataset tuple shape, train/test dataloader APIs, transforms, samplers, or preflight checks before training/evaluation.

Start here

  1. Identify the dataset name exactly as FastReID expects it, for example Market1501, DukeMTMC, MSMT17, VeRi, VehicleID, SmallVehicleID, VeRiWild, or SmallVeRiWild.
  2. Resolve the dataset root that FastReID will receive: FASTREID_DATASETS if set, otherwise a relative datasets directory from the process working directory.
  3. Validate the tree before launching any train/eval job:
python sub-skills/data-and-datasets/scripts/validate_dataset_layout.py --root <datasets-root> --dataset Market1501
  1. If using a custom dataset, import the module that registers it before calling FastReID's train/test loader builders.

Runtime references and scripts

  • references/dataset-formats.md: use for built-in dataset names, expected directory/list-file layouts, file-name parsing rules, and the canonical (img_path, pid, camid) item schema.
  • references/data-api.md: use for DATASET_REGISTRY, ImageDataset, CommDataset, build_reid_train_loader, build_reid_test_loader, transform options, sampler choices, and custom dataset registration patterns.
  • references/troubleshooting.md: use when FastReID cannot find datasets, parses pid/camid incorrectly, query/gallery is empty, samplers fail because of batch settings, or old dataset tests/imports conflict with this API.
  • scripts/validate_dataset_layout.py: safe no-download validator for selected built-in layout roots; run it before expensive train/eval commands or when diagnosing partial dataset trees.

Route boundaries

  • Training loops, schedulers, checkpoints, distributed launch, evaluator output, and train/eval command construction belong to ../training-and-evaluation/.
  • Model registry, backbone/head/meta-architecture construction, predictors, feature tensors, and checkpoint loading belong to ../modeling-and-inference/.
  • Source-only package setup, config inheritance, YAML merge behavior, and config recipe selection belong to ../setup-and-configuration/.
  • Project-specific datasets under FastReID extension projects belong to ../deployment-and-projects/; use this sub-skill only for the shared registry/layout mechanics they rely on.

Verified operating facts

  • FASTREID_DATASETS defaults to datasets when the environment variable is unset.
  • Dataset records are tuples (img_path, pid, camid). Training built-ins usually prefix training pid/camid with a dataset name string before CommDataset(..., relabel=True) maps them to contiguous integers.
  • Built-in registry keys include person ReID datasets (Market1501, DukeMTMC, MSMT17, CUHK03, VIPeR, GRID, iLIDS, etc.) and vehicle ReID datasets (VeRi, VehicleID, SmallVehicleID, MediumVehicleID, LargeVehicleID, VeRiWild, SmallVeRiWild, MediumVeRiWild, LargeVeRiWild).
  • API signatures verified for this FastReID version include build_reid_train_loader(train_set, *, sampler=None, total_batch_size, num_workers=0) and build_reid_test_loader(test_set, test_batch_size, num_query, num_workers=4); the config system also permits calls such as build_reid_train_loader(cfg, combineall=...) and build_reid_test_loader(cfg, dataset_name=...).

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351