---

name: evaluation-and-visualization description: "Evaluate rPPG and BigSmall outputs, inspect saved signals and preprocessed arrays, and choose portable motion-analysis visualizations." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Evaluation and visualization

Use this sub-skill after a model or unsupervised method has produced signals. It turns saved predictions into HR/RR metrics, compares traces, inspects cached arrays, or summarizes OpenFace motion. It does not select or run a model, build data caches, or reproduce training.

Choose a path

1. Metrics: use the repository evaluator when the checkout and its config are available. Preserve LABEL_TYPE, test FS, evaluation method, and window settings. The exact contracts and known implementation edge cases are in evaluation-api.md.
2. Saved signal inspection: use scripts/plot_saved_predictions.py for a noninteractive PNG/SVG/PDF plot. The pickle must contain predictions, labels, label_type, and fs; see visualization.md.
3. Preprocessed-array inspection: use scripts/plot_preprocessed_arrays.py for a static frame plus label time/frequency plot. It is the portable replacement for the interactive notebook; it accepts one input .npy and its matching label .npy.
4. Motion: if MP4s are needed, use scripts/convert_frames_to_mp4.py, then run an already installed OpenFace FeatureExtraction executable outside this skill. Summarize two OpenFace CSV directories with scripts/summarize_openface_motion.py.

Do not use a notebook, source-checkout import, bundled model, sample output, or vendor/OpenFace code as a runtime dependency. The original motion workflow needs OpenFace installed separately; it is not replaced by the converter.

Evaluation guardrails

• DiffNormalized labels/predictions set diff_flag=True: cumulatively sum before detrending. Raw and Standardized set it to False.
• Pass the test sampling rate (FS, commonly 30) rather than assuming it.
• The evaluator can use FFT or peak detection (spelling is case-sensitive in the repository config). FFT searches the configured band; peak detection uses peak spacing. A window shorter than 9 samples is ignored.
• The post-processing default band is 0.6--3.3 Hz; the source comments recommend 0.75--2.5 Hz for paper-style results. Keep the chosen band in the report. MACC is maximum absolute lagged correlation and SNR is reported in dB.
• Request MAE, RMSE, MAPE, Pearson, SNR, MACC, and/or BA as appropriate. Do not report undefined values from empty, constant, or too-short arrays as successful evaluation; follow the failure guidance in troubleshooting.md.

Output locations

Test output pickles are normally under the configured log path at the test experiment's saved_test_outputs directory. Bland--Altman files are normally under the log path, experiment name, and bland_altman_plots. Exact filenames, including FFT/Peak variants, are listed in evaluation-api.md. Static helper outputs go only where --output names them; existing files are never overwritten unless --force is explicitly supplied.

Cross-skill boundaries

• Use supervised-models for producing neural predictions and understanding trainer output timing.
• Use data-preparation for cache filenames, preprocessing schema, and input/label alignment.
• Use unsupervised-methods for selecting signal-processing algorithms before evaluation.
• Keep this sub-skill focused on evaluation, inspection, plotting, and motion; do not duplicate those setup workflows.

Portable commands

python scripts/plot_saved_predictions.py --help
python scripts/plot_saved_predictions.py --input outputs.pickle --output signal.png --trial 0 --chunk-size 180 --chunk 0
python scripts/plot_preprocessed_arrays.py --input 000_input0.npy --label 000_label0.npy --output cache.png
python scripts/convert_frames_to_mp4.py --help
python scripts/convert_frames_to_mp4.py --mode ubfc-rppg --input-dir <dataset-root>/UBFC --output-dir <scratch-dir>/ubfc-mp4
python scripts/summarize_openface_motion.py --input-dir <scratch-dir>/openface-a --compare-dir <scratch-dir>/openface-b --output motion.png

All commands are local-only and noninteractive. Read the linked references before changing bands, label transforms, AU definitions, or output paths.