Simple GAN with PyTorch

This repository contains a basic implementation of a Generative Adversarial Network (GAN) using PyTorch, designed to generate handwritten digits similar to the MNIST dataset.

---

Overview

• Generator (G): Takes random noise as input and tries to produce realistic-looking MNIST digit images.
• Discriminator (D): Attempts to distinguish between real MNIST images and the fake images generated by the Generator.
• Both models are trained together in a zero-sum game — G tries to fool D, while D tries not to be fooled.

---

Files

• GAN.ipynb — Colab notebook with the full PyTorch code for:

  • Loading the MNIST dataset.
  • Defining the Generator and Discriminator networks.
  • Training loop with logging to TensorBoard.

---

Requirements

• Python 3
• PyTorch
• Torchvision
• TensorBoard

You can install dependencies using:

pip install torch torchvision tensorboard

---

How to Run

1. Open in Colab:

2. Run the notebook cells:

   • The notebook will download MNIST automatically.
   • The GAN will train for 50 epochs (by default).

3. Monitor progress:

   • Generated fake images and real samples are logged to TensorBoard.
   • Run tensorboard --logdir=logs to see them.

---

Training Details

• Latent Vector (z-dim): 64
• Image Dimension: 28×28 = 784
• Optimizer: Adam (lr = 3e-4)
• Loss: Binary Cross Entropy (BCE)

---

Output

• During training, you can visually compare real vs. generated images.
• Over time, fake images should look more realistic.

---