ORAGEN

Goal

Run the real ORAGEN example for audio-visual occlusion-robust gender recognition and age estimation, including video inference.

What you will learn

This example combines plugin models with a registry-driven inference pipeline, remote/local checkpoint resolution, and optional DAG execution.

Prerequisites

Install the plugin and its dependencies:

python -m pip install -e examples/oragen

Training requires ORAGEN-style cached features, labels, VAD files, corpora, and the paths in the selected config. Inference additionally requires a readable video, example dependencies, a torchcodec runtime or ffmpeg fallback, network access on first checkpoint download (or local/cached weights), and writable work space. Ultralytics resolves yolo26n.pt separately.

Run inference

chimera-ml inference \
  --input examples/oragen/samples/f_25.mp4 \
  --output prediction.json \
  --config-path examples/oragen/configs/inference.yaml

The pipeline extracts mono 16 kHz audio, runs Silero VAD, samples video, detects a face, builds fixed windows, resolves audio/image/fusion checkpoints, extracts features, fuses predictions, and aggregates windows. The output step writes JSON with input and predictions.

Run training/evaluation

chimera-ml validate-config -c examples/oragen/configs/multimodal_train.yaml
chimera-ml train -c examples/oragen/configs/multimodal_train.yaml
chimera-ml eval -c examples/oragen/configs/multimodal_test.yaml --checkpoint-path path/to/checkpoint.pt

Audio configs use the same commands with audio_train.yaml and audio_test.yaml.

The authoritative paths, component keys, and model preparation steps remain in the ORAGEN example.