😊 Emotion Classifier — Custom CNN + Streamlit

A real-time facial emotion classification app built with a custom CNN trained from scratch in PyTorch. The Streamlit web app supports both image upload and live webcam inference via WebRTC.

---

📸 Demo

App predicting sad with 68.1% confidence, showing full probability breakdown across all 7 emotion classes

---

🧠 Model Architecture

A custom 3-block CNN built in PyTorch — no pretrained backbone, trained from scratch.

Input (1×128×128 grayscale)
    │
    ├── Block 1: Conv2d(1→32)   → BN → ReLU → Conv2d → BN → ReLU → MaxPool → Dropout(0.1)
    ├── Block 2: Conv2d(32→64)  → BN → ReLU → Conv2d → BN → ReLU → MaxPool → Dropout(0.1)
    ├── Block 3: Conv2d(64→128) → BN → ReLU → Conv2d → BN → ReLU → MaxPool → Dropout(0.1)
    │
    ├── AdaptiveAvgPool2d(1×1)
    └── Classifier: Linear(128) → ReLU → Dropout(0.5) → Linear(7)

Key design choices:

• Grayscale input (1 channel) — removes color bias, focuses on facial structure
• BatchNorm after every conv layer — faster convergence and stable training
• AdaptiveAvgPool — makes the model input-size flexible
• Dropout at 0.1 (conv blocks) and 0.5 (FC layer) — reduces overfitting

---

📊 Model Performance

Trained and evaluated on the Facial Emotion Recognition dataset (Kaggle) — 7,470 test samples.

Emotion	Precision	Recall	F1-Score	Support
Angry	0.53	0.40	0.46	888
Disgust	0.77	0.77	0.77	888
Fear	0.47	0.33	0.39	888
Happy	0.86	0.80	0.83	1711
Neutral	0.57	0.64	0.60	1226
Sad	0.43	0.65	0.52	981
Surprise	0.76	0.70	0.73	888
Overall Accuracy			63%	7470

Happy and Disgust are the strongest classes. Fear and Angry are the hardest to classify — consistent with findings in the literature on this dataset due to visual similarity between emotions.

---

🏷️ Emotion Classes

angry · disgust · fear · happy · neutral · sad · surprise

---

🗂️ Project Structure

emotion-classifier-cnn/
├── app.py                  # Main Streamlit app (image upload + webcam tabs)
├── script.py               # Image upload inference module
├── video.py                # Webcam real-time inference module
├── best_emotion_cnn.pt     # Trained model weights
├── class_names.json        # Emotion label list
├── requirements.txt        # Python dependencies
└── README.md

---

⚙️ Setup & Installation

1. Clone the repository

git clone https://github.com/Alnurabda/emotion-classifier-cnn.git
cd emotion-classifier-cnn

2. Create a virtual environment

python -m venv venv
source venv/bin/activate        # macOS/Linux
venv\Scripts\activate           # Windows

3. Install dependencies

pip install -r requirements.txt

4. Run the app

streamlit run app.py

Open your browser at http://localhost:8501

---

🚀 Features

• Upload Image tab — drag & drop any face image and get:
  • Top predicted emotion with confidence score
  • Full probability bar chart across all 7 classes
  • Raw probability table
• RealTime tab — live webcam feed with:
  • Per-frame emotion prediction overlaid on video
  • Optional moving-average smoothing to stabilize predictions
• Runs entirely on CPU — no GPU required

---

🔧 Configuration

IMG_SIZE = 128   # Input resolution (must match training)
IN_CH    = 1     # 1 = grayscale, 3 = RGB
TOPK     = 3     # Top-K predictions shown
SMOOTH_N = 8     # Webcam smoothing window (frames)

---

📦 Dependencies

Package	Purpose
torch / torchvision	Model definition & inference
streamlit	Web UI
streamlit-webrtc	Real-time webcam streaming
opencv-python	Frame processing
Pillow	Image loading
numpy / pandas	Data handling & display
av	Video frame decoding

---

🙋 Author

Built by Alnura · GitHub