---

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