AGENT SKILL

training-evaluation-scripts

Construct safe GluonCV training, evaluation, demo, and benchmark flag templates from the script zoo without executing source checkout scripts.

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

01 前言 Front Matter

name
training-evaluation-scripts
license
Apache 2.0
metadata
{ "package": "gluoncv", "disco-role": "operating", "parent-skill": "gluon-cv" }
description
Construct safe GluonCV training, evaluation, demo, and benchmark flag templates from the script zoo without executing source checkout scripts.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

GluonCV training and evaluation script workflows

Use this sub-skill when a task asks for GluonCV training, evaluation, demo, feature extraction, benchmarking, DDP/Horovod/DALI usage, or command-line flags for the script zoo: CIFAR/ImageNet classification, SSD/YOLO/Faster R-CNN/CenterNet/Mask R-CNN detection, segmentation, pose, action recognition, Monodepth2, SiamRPN/SMOT tracking, GAN, Re-ID, dataset preparation, or AutoGluon example commands.

This sub-skill turns the repository's script families into self-contained, non-executable flag templates and checklists. Treat original programs as evidence for flag shape, not as runtime files this skill depends on.

Route first

Stay here for:

  • Choosing a script family for training/evaluation/demo/benchmarking.
  • Building a flag template and identifying required dataset/model/backend inputs.
  • Explaining common flags such as --model, --network, --dataset, --dataset-root, --data-dir, --gpus, --num-gpus, --batch-size, --num-workers, --epochs, --resume, --pretrained, --deploy, --quantized, and --config-file.
  • Deciding whether a workflow is safe to dry-run, help-only, reference-only, or requires user approval because it downloads, trains, benchmarks, or needs GPUs.

Route elsewhere when the user needs deeper API or data details:

  • MXNet model names, get_model, reset_class, and inference API details: ../mxnet-model-zoo/
  • Dataset layouts, transforms, loaders, and annotation validation: ../data-transforms-datasets/
  • PyTorch action-recognition models, configs, tensor shapes, and DDP model smokes: ../torch-video-workflows/
  • AutoGluon tasks, export, ONNX, TVM, quantized deployment: ../automl-deployment-export/

Safe workflow

  1. Identify the task family and backend: MXNet script-style workflow, Torch action-recognition config workflow, AutoGluon example, deployment/export, or dataset preparation.

  2. Read references/script-catalog.md to classify source families as adapted, reference-only, or excluded.

  3. Use scripts/build_training_command.py to generate a no-side-effect flag template:

    python scripts/build_training_command.py detection-yolo \
      --dataset voc --dataset-root /data/VOCdevkit --gpus "" --batch-size 2
    
  4. Replace placeholders with user-provided values and verify prerequisites before any real execution in a user-provided checkout or project: dataset root, annotations, image/video files, pretrained cache/network policy, backend framework, optional dependencies, and GPU availability.

  5. For CPU-safe exploration, prefer dry model/data smoke checks from sibling sub-skills before launching long jobs.

  6. For real training, explicitly confirm expected runtime, write locations, and resource usage. Long workflows can create checkpoints/logs, download weights/data, and consume GPUs for hours.

References and helper

  • references/training-and-evaluation-scripts.md — task-family flag anatomy and adaptation patterns.
  • references/script-catalog.md — source script family inventory, selected bundled replacement, and skip/risk decisions.
  • references/troubleshooting.md — dataset, backend, optional dependency, OOM, resume, distributed, and cache failures.
  • scripts/build_training_command.py — safe flag-template builder; it prints pseudo-command families and warnings but never starts training.

Safety defaults

  • Use empty --gpus "" or --num-gpus 0 for CPU templates when the underlying workflow supports it.
  • Do not run benchmark, DDP, Horovod, DALI, video decoding, dataset download, export, or training commands until the user confirms hardware/data/network/write side effects.
  • Treat pretrained downloads as network/cache side effects. Prefer pretrained=False in API smoke checks and pass explicit cache/write locations for real runs when possible.
  • When a source family defaults to GPU, make the GPU requirement visible and change the template only when CPU behavior is supported by that workflow.

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