😄 Facial Expression Generation Using GAN and Autoencoder

This project leverages deep learning to generate facial expressions from labeled data using an autoencoder architecture trained on the CelebA dataset. The model learns facial features (e.g., "Smiling") and can reconstruct or generate faces conditioned on these attributes.

🔍 Overview

Facial expression generation is a complex area in computer vision. This notebook uses the CelebA dataset, focusing on labeled facial attributes to generate or classify expressions using an autoencoder structure.

📁 Project Structure

Facial_Expression_Generation_Labeled.ipynb
README.md

🚀 Features

• 📂 Loads and filters CelebA facial attribute dataset
• 🧠 Trains a custom encoder-decoder (autoencoder) network
• 🎨 Generates expressions based on labeled features (e.g., "Smiling")
• 📊 Tracks and visualizes accuracy, loss, and outputs
• 🧪 Evaluates generation quality and classification accuracy

🧰 Technologies Used

• Python 3.x
• PyTorch
• torchvision
• Google Colab
• Matplotlib, Pandas
• PIL, NumPy

📊 Dataset

CelebA — a large-scale face dataset with 202,599 celebrity images, each annotated with 40 attribute labels (e.g., Smiling, Male, Eyeglasses).

This project focuses on the "Smiling" label to:

• Train the encoder to learn a latent representation of smiling vs non-smiling faces
• Use the decoder to generate corresponding facial images

🧪 Model Evaluation

The training process monitors:

• Reconstruction Loss (MSE or BCE): Measures how well the decoder recreates the input.
• Attribute Accuracy: A simple classifier is sometimes used on the latent representation to test if attributes like "Smiling" are preserved.
• Loss & Accuracy Visualization: The notebook includes plots to visualize training dynamics.

⚙️ How to Use

1. Clone the repo or download the notebook.
2. Upload the CelebA dataset and attribute file (img_align_celeba.zip and list_attr_celeba.txt) to your Google Drive.
3. Run the notebook on Google Colab (recommended).
4. All required packages will be installed at runtime.

📈 Output Examples

The notebook generates:

• Reconstructed face images side-by-side with originals

• Images with changed expression attributes

• Loss and accuracy graphs per epoch -Confusion Matrix -Hyperparamater Tuning

📌 Notes

• The model uses a custom CelebAZipDataset class to read images from a ZIP file without extracting.
• Ensure you have enough runtime memory in Colab (recommended: GPU backend).
• The model can be extended to other labels like “Happy,” “Sad,” or “Surprised” by changing attribute filters.

✅ Future Work

• Extend to full expression classification (multi-label)
• Use a Conditional GAN for sharper results
• Evaluate with metrics like FID or IS scores
• Add a web demo using Streamlit or Gradio

👨‍💻 Author

Sanya Shresta Jathanna