ramyamail2/Classroom-Stress-Detection

🧠 Face Emotion Recognition · Deep Learning Mini Project

A complete, from-scratch implementation of a Convolutional Neural Network for 7-class facial emotion recognition — no TensorFlow/PyTorch required.

📁 Project Structure

emotion_recognition/
│
├── model.py          # CNN layers, EmotionCNN class, FER dataset generator
├── train.py          # Full training pipeline + visualisation
├── inference.py      # Real-time webcam / image / demo inference
├── dataset_utils.py  # Data loading, augmentation, class weights
├── demo.html         # Self-contained interactive browser demo
├── requirements.txt  # Python dependencies
│
└── outputs/          # Generated after training
    ├── best_weights.npy
    ├── metrics.json
    ├── sample_images.png
    ├── class_distribution.png
    ├── architecture.png
    ├── training_curves.png
    ├── confusion_matrix.png
    └── per_class_metrics.png

🚀 Quick Start

1. Install dependencies

pip install -r requirements.txt

2. Quick smoke-test (3 epochs, synthetic data)

python train.py --quick

3. Full training on synthetic dataset (FER-2013 style)

python train.py

4. Use real FER-2013 dataset

# Download via Kaggle CLI
kaggle datasets download -d msambare/fer2013 -p data --unzip
python train.py --fer2013 data/fer2013.csv

5. Run inference

# Demo grid (no camera needed)
python inference.py --demo

# On a single image file
python inference.py --image path/to/face.jpg

# Live webcam (requires camera)
python inference.py --webcam

6. Interactive browser demo

Simply open demo.html in any modern browser — no server needed.

🏗️ Model Architecture

Input (1×48×48)
  ↓
Conv2D(32, 3×3) → BatchNorm → ReLU → MaxPool(2×2)   →  32×24×24
Conv2D(64, 3×3) → BatchNorm → ReLU → MaxPool(2×2)   →  64×12×12
Conv2D(128,3×3) → BatchNorm → ReLU → MaxPool(2×2)   → 128×6×6
  ↓
Flatten → 4608
Dense(256) → Dropout(0.5)
Dense(128) → Dropout(0.3)
Dense(7)   → Softmax

Parameters: ~1.2 million Framework: Pure NumPy (no TF/PyTorch)

📊 Dataset: FER-2013

Images: 48×48 grayscale Source: Kaggle – msambare/fer2013

📈 Expected Performance (Real FER-2013)

Note: The synthetic dataset used for quick testing produces low accuracy by design — it only simulates the data structure. Use real FER-2013 for production-quality results.

🔧 Implementation Notes

No External DL Framework

All layers are implemented from scratch in NumPy:

• —ConvLayer — 2-D convolution with He initialisation
• —BatchNorm — channel-wise batch normalisation
• —MaxPool2x2 — 2×2 max pooling with correct backprop
• —Dropout — inverted dropout
• —Dense — fully-connected layer
• —Softmax — numerically stable softmax + CE loss

Augmentation (dataset_utils.py)

• —Horizontal flip
• —Random rotation (±12°)
• —Translation (±4 px)
• —Zoom (0.9–1.1×)
• —Brightness jitter
• —Gaussian noise

Class Imbalance

compute_class_weights() returns per-class weights for weighted cross-entropy, especially important for the underrepresented Disgust class (436 samples).

📄 License

MIT — free for academic and personal use.

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference