README.md

---
title: Totes Emosh
emoji: ⚡
colorFrom: yellow
colorTo: indigo
sdk: gradio
sdk_version: 6.15.2
app_file: app.py
pinned: false
short_description: Pose six basic emotions. See what the classifier reads.
---

# totes-emosh

LittleMonkeyLab toy app for facial-expression recognition. A
**six-emotion replication challenge**: pose each of the six basic
emotions (happy, sad, fear, disgust, anger, surprise), the classifier
reads each attempt, and you leave with a single-page **EmotionMap**
PDF.

## What this does

Companion app to Week 3 Part 4 of the Goldsmiths MSc in Psychology —
*Emotion in Action*. Built around a single-person replication of
Porter and ten Brinke (2008): can you fake an emotional expression
on demand, and how well does a state-of-the-art classifier read each
attempt? Two output formats: **Face** (your face crops) and
**Wireframe** (anonymised landmark meshes — face-free).

Codewright fork of the original
[`LittleMonkeyLab/All_a_bit_emotional`](https://huggingface.co/spaces/LittleMonkeyLab/All_a_bit_emotional)
Space, slimmed to static images only. The dynamic video pipeline now
lives elsewhere.

## Setup

```bash
git clone <url>
cd totes-emosh
uv venv
uv pip install -r requirements.txt
uv run python app.py
```

`requirements.txt` is preserved for HuggingFace Space deployment.
`pyproject.toml` is the codewright-canonical dependency manifest for local
development.

## Deployment

This repo is designed to deploy directly as a HuggingFace Space (the
YAML frontmatter above is the Space config). The face landmarker and
classifier use the modern `mediapipe.tasks` API, which runs on both
Apple Silicon (local dev) and the Linux x86-64 Spaces runtime
unchanged.

### Model assets

No model weights are committed to this repo. All weights download on
first request:

- `FER_static_ResNet50_AffectNet.pt` from HuggingFace
  (`ElenaRyumina/face_emotion_recognition`), driven by `config.toml`
  and `app/model.py`.
- `face_landmarker.task` (MediaPipe Face Landmarker bundle, ~3 MB)
  from `storage.googleapis.com/mediapipe-models/` on first inference.

Cold-start latency from these downloads is a few seconds on the
ResNet50 weights and trivial on the landmarker bundle.

### Runtime expectations

CPU inference on a free HF Space (~2 vCPU, 16 GB RAM):

- MediaPipe Face Landmarker: ~10 ms / frame
- ResNet50 static classifier: ~50–100 ms / frame

## Status

Active. Original `LittleMonkeyLab/All_a_bit_emotional` Space remains
live and unchanged; this is the new canonical six-emotion build under
its own slug (`LittleMonkeyLab/totes-emosh`).

## Provenance

Cloned from `LittleMonkeyLab/All_a_bit_emotional` via `hf download` on
2026-06-03; substantially rebuilt 2026-06-05/07. See `DEVLOG.md` for
the iteration trail.

## Credit

Created by Dr. Gordon Wright — A LittleMonkeyLab caper.
Part of the Goldsmiths MSc in Psychology, Week 3 Part 4.