AGENT SKILL

3d-resnets-pytorch

Routes 3D ResNets PyTorch video action-recognition workflows across training, inference, and data preparation.

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

01 前言 Front Matter

name
3d-resnets-pytorch
license
MIT
metadata
{ "disco-role": "operating" }
description
Routes 3D ResNets PyTorch video action-recognition workflows across training, inference, and data preparation.
disable-model-invocation
true

02 资源文件 8

03 SKILL.md 正文

3D ResNets PyTorch

Use this root skill when the user asks about the 3D ResNets PyTorch repository, its CLI flags, its dataset layouts, or its video action-recognition workflows.

Read first

  • references/cli-reference.md
  • references/troubleshooting.md
  • references/repo-provenance.md
  • scripts/check_imports.py
  • scripts/check_main_help.py

What this skill covers

  • Training, validation, inference, and fine-tuning for video action recognition.
  • Dataset preparation from raw videos into JPEG frame trees, RGB HDF5 files, and annotation JSONs.
  • Checkpoint handling, model-family selection, and result evaluation.
  • Common data-layout and runtime compatibility pitfalls.

Route to a sub-skill

training-and-inference

Use this route for:

  • Fresh training runs, resume flows, and pretrained fine-tuning.
  • Validation-only runs and sliding-window inference.
  • Result scoring and DataParallel checkpoint cleanup.
  • Questions about model families, class counts, ft_begin_module, or resume_path / pretrain_path behavior.

Read:

  • sub-skills/training-and-inference/SKILL.md
  • sub-skills/training-and-inference/references/workflows.md
  • sub-skills/training-and-inference/references/model-catalog.md
  • sub-skills/training-and-inference/references/troubleshooting.md
  • sub-skills/training-and-inference/scripts/evaluate_results.py
  • sub-skills/training-and-inference/scripts/strip_dataparallel.py

data-preparation

Use this route for:

  • Extracting JPEG frames or RGB HDF5 files from raw videos.
  • Building Kinetics, UCF101, HMDB51, MIT, or ActivityNet JSON metadata.
  • Adding ActivityNet fps fields.
  • Questions about class directories, split files, HDF5 manifests, or jpg versus hdf5 layout.

Read:

  • sub-skills/data-preparation/SKILL.md
  • sub-skills/data-preparation/references/workflows.md
  • sub-skills/data-preparation/references/data-formats.md
  • sub-skills/data-preparation/references/troubleshooting.md
  • sub-skills/data-preparation/scripts/extract_video_frames.py
  • sub-skills/data-preparation/scripts/extract_video_hdf5.py
  • sub-skills/data-preparation/scripts/build_annotation_json.py

Quick runtime helpers

  • scripts/check_imports.py verifies that the core source modules import from a checkout and applies the temporary legacy Scale alias when needed.
  • scripts/check_main_help.py prints the full main.py CLI help through the same compatibility shim.
  • scripts/run_main.py forwards into the repository CLI after preparing the checkout and compatibility shim.

Shared environment facts

This repo expects a Python environment with PyTorch, torchvision, pandas, h5py, scikit-learn, joblib, and FFmpeg/FFprobe available on PATH.

A modern torchvision wheel may not expose torchvision.transforms.Scale. If that happens, use the compatibility shim in scripts/_torchvision_compat.py or a legacy torchvision release that still ships Scale.

Route selection guidance

  • If the user already has prepared videos and annotation JSONs, start with training-and-inference.
  • If the user still needs frames, HDF5 files, or split JSONs, start with data-preparation.
  • If the request mentions both, do data preparation first unless the videos and labels are already ready.
  • If the user only wants command discovery or environment checks, use the root helpers and the sub-skill references rather than reopening the source repo.

Common handoff sequence

  1. Prepare or verify the dataset layout with data-preparation.
  2. Verify the environment with scripts/check_imports.py.
  3. Inspect the CLI with scripts/check_main_help.py.
  4. Run the desired training or inference command with scripts/run_main.py.
  5. Score or clean outputs with the training-and-inference helpers.

Don’t do this

  • Do not point future agents back to the original checkout paths.
  • Do not use flow with JPEG inputs.
  • Do not assume resume checkpoints can change architecture.
  • Do not score --inference_no_average JSON with the result evaluator before aggregating segment outputs.

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