AGENT SKILL

training-and-inference

Routes training, fine-tuning, validation, checkpoint, and inference workflows for 3D ResNets PyTorch.

License 许可
MIT
最近更新
2026-10-08
Stars
★ 11
Forks
3

01 前言 Front Matter

name
training-and-inference
license
MIT
metadata
{ "disco-role": "operating" }
description
Routes training, fine-tuning, validation, checkpoint, and inference workflows for 3D ResNets PyTorch.
disable-model-invocation
true

02 资源文件 3

03 SKILL.md 正文

training-and-inference

Use this sub-skill when you need to train, fine-tune, validate, resume, or run inference for this repository's main.py-style workflows without reopening the source checkout.

Use this route for

  • Fresh training runs and stage-wise fine-tuning.
  • Checkpoint save/resume, optimizer state recovery, and milestone handling.
  • Loading pretrained weights with a different final class count.
  • Validation, multi-clip inference, and result scoring.
  • DataParallel checkpoint cleanup before reuse or inspection.

Do not use this route for

  • Raw video extraction or annotation JSON creation. Use data-preparation.
  • Repo-wide routing or root skill selection. Use the root router.
  • Model conversion or export outside the main.py training loop.

Read first

Skill-owned scripts

  • scripts/evaluate_results.py — score top-k recognition results against a ground-truth JSON.
  • scripts/strip_dataparallel.py — remove module. prefixes from checkpoint state dicts.

Typical flow

  1. Confirm the dataset tree and annotation JSON already exist.
  2. Pick a model family and depth from references/model-catalog.md.
  3. Choose scratch train, resume, or fine-tune flags in references/workflows.md.
  4. Run the root-level main.py wrapper from the repo skill tree.
  5. Evaluate results with scripts/evaluate_results.py or clean checkpoints with scripts/strip_dataparallel.py when needed.

Cross-links

04 同仓库更多 Skills