🧠 DCGAN: Generate Realistic Human Faces using PyTorch

This project uses Deep Convolutional GAN (DCGAN) architecture implemented in PyTorch to generate realistic human face images from a noise vector. It includes full training and image generation pipeline.

🔧 Features

Uses torchvision.datasets.ImageFolder to load real face images.
Implements a DCGAN:
- Generator using ConvTranspose2d layers.
- Discriminator using Conv2d layers without Sigmoid.
Trains on 64x64 images.
Supports CUDA (GPU) acceleration.
Saves sample images every 10 epochs.
Includes function to generate and display one random face at the end.

📁 Folder Structure

project/
│
├── images/ # Folder with real training face images
│ ├── class_x/ # ImageFolder requires class subdirectories
│ ├── image1.jpg
│ ├── image2.jpg
│
├── generated_epoch_10.png # Sample outputs saved every 10 epochs
├── generator.pth # Saved Generator weights after training
├── gan_faces.py # Python script with model and training code
└── README.md # This file

🚀 Getting Started

✅ Prerequisites

Install Python and required libraries:

pip install torch torchvision matplotlib tqdm

Make sure you have a folder ./images/ structured like:

images/
└── faces/
    ├── img1.jpg
    ├── img2.jpg

Where faces/ is any class name (required by ImageFolder).

🧠 How It Works

Generator: Takes a random noise vector and generates a 64x64x3 image.
Discriminator: Takes a real or fake image and returns a score (real or fake).
Loss Function: BCEWithLogitsLoss is used for both Generator and Discriminator.
Label Smoothing: Real images use soft labels (0.9 instead of 1.0) to stabilize training.

⚙️ Training

To start training:

python gan_faces.py

The script will:

Load images from ./images/
Train for 100 epochs (can be changed)
Save sample images every 10 epochs as generated_epoch_X.png
Save final model as generator.pth

🧪 Generate One Face

At the end of training, a sample face will be displayed using matplotlib.

You can also use this standalone function:

generate_one_face("generator.pth")

To generate and visualize a face anytime.

ankitkushwaha90
/

DCGAN