AGENT SKILL

training

Routes Ultra-Fast-Lane-Detection training, checkpoint, resume, finetune, and multi-GPU launch workflows.

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

01 前言 Front Matter

name
training
license
MIT
metadata
{ "disco-role": "operating" }
description
Routes Ultra-Fast-Lane-Detection training, checkpoint, resume, finetune, and multi-GPU launch workflows.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

training

Use this sub-skill when a task is about fitting the lane detector: choosing optimizer and scheduler settings, launching single-GPU or distributed training, resuming checkpoints, or understanding how the repo logs and saves models.

Read this when

  • You need to start a training run from train.py.
  • You need to resume or finetune from a checkpoint.
  • You need to translate the repo's shell launcher into a safer command.
  • You need to explain the repo's losses, metrics, or checkpoint naming.

What this sub-skill owns

  • Training command construction.
  • DDP/NCCL launch advice.
  • Optimizer, scheduler, loss, and metric selection.
  • Checkpoint save/resume/finetune behavior.
  • TensorBoard logging and code-backup side effects.

What it does not own

  • Dataset root preparation and TuSimple conversion: see data-and-config.
  • Evaluation, scoring, and demo output files: see evaluation-and-visualization.
  • TorchScript export and benchmark timing: see export-and-speed.

Start here

  • Read references/training-workflows.md for the practical command patterns.
  • Read references/api-reference.md for verified function signatures and model dimensions.
  • Read references/troubleshooting.md for CUDA, checkpoint, and launch failures.
  • Run scripts/model_cuda_smoke.py when you need a quick proof that the model can build and run on CUDA.
  • Use scripts/launch_training_template.sh as the safer replacement for the repo's raw launch shell snippet.

Typical flow

  1. Pick the dataset family and matching config from data-and-config.
  2. Decide whether the run is single-GPU or distributed.
  3. Set data_root, log_path, and any checkpoint path overrides explicitly.
  4. Confirm the backbone, griding_num, num_lanes, and use_aux match the chosen dataset family.
  5. Start the run and keep the logs outside the repository tree.

Caution points

  • The source training loop calls .cuda() directly.
  • The launcher assumes NCCL for multi-GPU runs.
  • cp_projects can copy a large working tree if log_path points at a directory inside the repo.
  • The checkpoint filename convention matters for resume logic.

Reference and script links

  • references/training-workflows.md for the command patterns and checkpoint lifecycle.
  • references/api-reference.md for the verified signatures.
  • references/troubleshooting.md for launch, CUDA, and checkpoint problems.
  • scripts/launch_training_template.sh for a parameterized shell launcher.
  • scripts/model_cuda_smoke.py for a small model/build smoke check.

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