Five modalities.
One interpretation.
Omni-CNN is SensAffect’s missing-modality-aware affect model. MMRY Cloud ships the zip and accepts the JSON. This host does not run PyTorch, does not treat a face or a voice as a diagnosis, and does not use the 128-d embedding as recall.
1 / FIVE GATED BRANCHES
Any sensor may be
absent.
At least one branch is required. No branch is ground truth. Behaviour is derived features only — never raw keystrokes.
- 01VisualB×3×112×112 aligned RGB crop — face, posture or scene.
- 02VocalB×80×T log-mel spectrogram. Prosody, not a transcript by default.
- 03LanguageB×L local token IDs. Zero is padding.
- 04PhysiologicalB×8×T heart, EDA, respiration or motion — consented, aligned.
- 05BehaviourB×16×T latency, dwell, correction, cadence and kin. See behavior_schema.json.
2 / IMPRINT JSON
Feeling beside memory.
Never instead of it.
Run infer.py. You get emotion (12 classes), VAD, intensity, uncertainty, per-modality weights, and a 128-d embedding. Paste that object into the workspace Omni-CNN tab, or POST it as imprint.
- 01Write the memory in your wordsThat row is the Memory side. Omni-CNN must not overwrite it.
- 02Attach the imprintStored as feeling_interpretations (derived) plus metadata.imprint.
- 03Do not search the feelingRecall is title and content. Asking “calm” abstains unless the word is in the record.
Expected result A citeable memory. A labelled interpretation with method sensaffect-omni-cnn. An embedding that sits as provenance and is not an index. Empty feeling if you skip the JSON.
3 / QUICK START
Train locally.
This host only stores.
Intel macOS: Python 3.12. Pin NumPy below 2. The included trainer is synthetic — it proves the pipeline, not a person.
- 01Unzip and venv
python3.12 -m venv .venv && pip install -r requirements.txt - 02Train one epoch
python train.py --epochs 1 --output sensaffect_omni_cnn.pt - 03Infer JSON
python infer.py sensaffect_omni_cnn.pt - 04Optional ONNX
python export_onnx.py sensaffect_omni_cnn.pt— Hailo compile is separate.
4 / AUTHORITY
Current consent.
Not a remembered grant.
Capability sa_omni_cnn lives in the SensAffect sensor library. affect_analysis is the permission to derive a Feeling interpretation. Both are append-only allow or deny. Clinical certification is out of scope. Weighting and calibration remain unestablished.