Skip to content
Back to skills

Inference And Demos

ASecurity

"Use MMAction2 inference APIs, inferencers, labels, visualization

  • 247 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 8, 2026
devopstestingapi

Works with

  • api

Security analysis

A100/100

Pro scans all 5 files and shows the line behind each finding

Scanned September 8, 2026

npx -y skills add VectorSpaceLab/AREX-Skill --skill inference-and-demos --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Inference And Demos?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Inference And Demos
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/vectorspacelab-inference-and-demos/badge)](https://www.skillsdirectory.com/skills/vectorspacelab-inference-and-demos)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: inference-and-demos
description: "Use MMAction2 inference APIs, inferencers, labels, visualization
  outputs, and optional pose/detection demo flows."
disable-model-invocation: true
metadata:
  disco-role: operating
license: Apache 2.0
---

# MMAction2 inference and demos

Use this sub-skill when the user needs MMAction2 action-recognition inference, high-level inferencer usage, label mapping, prediction dumps, headless visualization, or the optional human-detection/pose paths used by skeleton and spatio-temporal detection workflows.

## Route here for

- Direct recognizer inference with `init_recognizer` and `inference_recognizer`, including access to `ActionDataSample.pred_score`.
- High-level `ActionRecogInferencer` or `MMAction2Inferencer` workflows using local configs/checkpoints, model aliases, label files, `pred_out_file`, and `vid_out_dir`.
- Video-file, rawframe-folder, decoded-array, audio-feature `.npy`, and already-packed pose/dict inputs when the selected API supports that input form.
- CPU/GPU device choice, safe build-only checks, and troubleshooting of decode, visualization, optional dependency, and config/checkpoint mismatch failures.
- Optional skeleton recognition and spatio-temporal detection flows that depend on separately installed detector/pose packages and local checkpoints.

## Route elsewhere

- Dataset annotation formats, data preparation, config authoring, and `--cfg-options` design: `../data-and-configs/SKILL.md`.
- Training, testing, evaluation metrics, distributed launch, and work directory management: `../training-and-evaluation/SKILL.md`.
- Model-family selection beyond inference, custom modules, registry extension, export, and deployment: `../models-and-extension/SKILL.md`.

## Choose the smallest safe entry point

1. **Build/config smoke only**: run [`scripts/mmaction2_inference_smoke.py`](scripts/mmaction2_inference_smoke.py) with `--config CONFIG.py --check-build-only`. It imports MMAction2, parses a local config, builds with `checkpoint=None`, and never downloads weights.
2. **Need raw scores or an `ActionDataSample`**: use `init_recognizer(..., device="cpu")` plus `inference_recognizer(...)`; read `result.pred_score` and map class indices with a user-supplied label file.
3. **Need packaged prediction/visualization output**: use `ActionRecogInferencer` for action recognition, especially when selecting `input_format="video"`, `"rawframes"`, or `"array"`.
4. **Need a unified demo-like wrapper**: use `MMAction2Inferencer(rec=..., rec_weights=..., device=..., label_file=...)`; use it one input at a time when saving visualizations.
5. **Need skeleton or spatio-temporal detection**: first confirm optional detector/pose dependencies and local detector/pose/action checkpoints; then follow the staged workflows in the references.

## Safety defaults

- Prefer `device="cpu"` unless the user explicitly asks for CUDA and the runtime proves CUDA is available. Several public APIs default to `"cuda:0"`, so pass the device explicitly on CPU-only hosts.
- Prefer local config and checkpoint paths. Model aliases or full model names can trigger metadata weight lookup; do not rely on remote downloads for smoke checks.
- Keep GUI display disabled on headless servers (`show=False`); save visualizations with `vid_out_dir` or request `return_vis=True` in code.
- When dumping predictions, use a file extension supported by MMEngine serialization and keep `return_datasamples=False` unless the consumer explicitly needs `ActionDataSample` objects.

## Local references

- [`references/api-reference.md`](references/api-reference.md) — verified signatures, defaults, input forms, result objects, devices, labels, and checkpoint/model selection.
- [`references/demo-workflows.md`](references/demo-workflows.md) — concrete API snippets, demo-like command patterns, visualization, skeleton, and spatio-temporal detection workflows.
- [`references/troubleshooting.md`](references/troubleshooting.md) — common inference failures and recovery steps, including optional dependency and video decode errors.
- [`scripts/mmaction2_inference_smoke.py`](scripts/mmaction2_inference_smoke.py) — safe build/inference smoke helper with signature printing and no downloads by default.

Files in this skill

  • SKILL.md4.2 KB
  • references/api-reference.md9.3 KB
  • references/demo-workflows.md9.9 KB
  • references/troubleshooting.md10.5 KB
  • scripts/mmaction2_inference_smoke.py7.9 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.