DCGAN-for-Face-Image-Generation

Overview

In this assignment, I implemented and trained a Deep Convolutional Generative Adversarial Network (DCGAN) using PyTorch to generate realistic facial images based on the CelebA-HQ 256x256 dataset. The project covers the full deep learning pipeline—from dataset preprocessing to evaluating the generated images using the Fréchet Inception Distance (FID) metric.

Dataset

• URL: Kaggle - CelebA-HQ Resized 256x256

Major Steps

1. Library Installation & Environment Setup

• Installed libraries: torch, torchvision, matplotlib
• Mounted Google Drive to access/save datasets and models

2. Dataset Handling

• Loaded the CelebA-HQ 256x256 dataset from Google Drive
• Preprocessing:
  • Resized and center-cropped images
  • Converted images to tensors and normalized to [-1, 1]
• Created a custom PyTorch Dataset and DataLoader with batch size 128

3. DCGAN Architecture

• Implemented Generator and Discriminator based on the DCGAN paper
• Generator:
  • Used ConvTranspose2d layers for upsampling
  • Applied BatchNorm2d and ReLU activations
• Discriminator:
  • Used Conv2d layers for downsampling
  • Applied LeakyReLU and BatchNorm2d
• Weight Initialization: Normal distribution

4. Training Setup

• Loss Function: Binary Cross Entropy (BCE)
• Optimizer: Adam (used for both Generator and Discriminator)
• Training Schedule:
  • Initial training: 5 epochs
  • Extended training: 150 epochs total
• Metrics tracked:
  • Generator & Discriminator losses
  • D(x): Discriminator output on real images
  • D(G(z)): Discriminator output on fake images

5. Image Generation

• Used a fixed noise vector to generate consistent image samples
• Saved generated images in generated_fakes_1000/

6. FID Score Evaluation

• Installed pytorch-fid for evaluation
• Compared generated images with original CelebA dataset
• FID Score Achieved: 161.74

7. Saving and Testing

• Saved the trained Generator model to Google Drive
• Visualized generated outputs for quality check
• Confirmed sample diversity and dataset integrity

8. Results & Observations

• DCGAN effectively learned to generate faces with CelebA-like features
• Generated faces displayed reasonable structure and diversity
• FID Score could be improved further through:
  • Longer training
  • Hyperparameter optimization

---

Final Evaluation

Fréchet Inception Distance (FID): 161.74

Tools Used

• Google Colab
• PyTorch
• TorchVision
• pytorch-fid
• Matplotlib

Author

Priteesh Madhav Reddy Karra
Date: April 2025