🧠 Valence Regressor (MLP)

A lightweight MLP model trained to predict valence from frame-level emotion probabilities (7 basic emotions).
Optimized for low-emotional, imbalanced datasets such as research interviews or controlled lab experiments.

📌 Model Overview

Input: 7 emotion probabilities (Neutral, Happy, Sad, Angry, Surprised, Scared, Disgusted)
Output: Continuous valence score ∈ [-1, 1]
Architecture:
- Input → Linear(7 → 64) → ReLU
- Linear(64 → 32) → ReLU
- Linear(32 → 1) → Output

🏋️‍♀️ Training Details

Dataset: Normalized per-frame predictions from an FER model
Labels: Valence values derived from emotion vectors (not manually annotated)
Task: Regression
Epochs: 20
Optimizer: Adam
Loss: MSE (Mean Squared Error)

Note: Sequence dynamics were not used — model trained on static frame-level data.

📊 Evaluation

Set	MAE	RMSE	R²
Validation	≈ 0.018	0.033	0.9928
Test	—	0.00035	0.9922

⚠️ Notes & Limitations

Labels were generated from emotion distributions, not from ground-truth manual annotations.
Best suited for:
- Low-expressiveness environments (e.g. astronauts, patients, lab participants)
- Scenarios with few facial changes and class imbalance
Not ideal for:
- High-emotional, spontaneous data (e.g. movies, TikTok videos)
- Multimodal fusion without proper normalization

🧩 Use Cases

Affective computing
Human-computer interaction (HCI)
Emotion-aware agents
Cognitive workload estimation
Social robotics & soft biometrics

💡 Example Usage

import torch
from model import MLPRegressor  # Define as per training

model = MLPRegressor(input_size=7)
model.load_state_dict(torch.load("pytorch_model.bin"))
model.eval()

emotion_probs = torch.tensor([[0.01, 0.95, 0.02, 0.01, 0.005, 0.003, 0.001]])  # [batch_size, 7]
valence = model(emotion_probs)
print(valence.item())  # e.g. 0.93