This repository is a hands-on implementation of a Deep Convolutional GAN (DCGAN) using TensorFlow, trained on the MNIST dataset of handwritten digits.
It is designed for learning purposes, showing how a Generator and Discriminator can be trained adversarially to produce realistic-looking digit images.
Sample outputs from training over 5 epochs:
| Epoch | Discriminator Loss | Generator Loss | Sample Output |
|---|---|---|---|
| 1 | 0.6533 | 1.3067 |  |
| 2 | 1.1693 | 1.0325 |  |
| 3 | 1.1358 | 1.1644 |  |
| 4 | 0.9302 | 1.0952 |  |
| 5 | 0.9230 | 1.3392 |  |
Sample outputs from training over 5 epochs:
| Epoch | Discriminator Loss | Generator Loss | Sample Output |
|---|---|---|---|
| 1 | 1.3690 | 0.9053 |  |
| 2 | 1.3009 | 0.9084 |  |
| 3 | 1.3500 | 0.9443 |  |
| 4 | 1.3614 | 0.9069 |  |
| 5 | 1.3124 | 0.9217 |  |
Training required between 30–60 GB of RAM and took around 1–3 hours depending on system load and configuration.
On my run, it took ~62 minutes.
For cloud experiments, renting an EC2 high-memory instance at spot pricing would keep training costs around $1 per run.
-
OS (my setup): Windows 11 Pro
-
Hardware (my setup): My Almighty HP EliteBook 840 G3
- Intel Core i5 (6th Gen)
- 16 GB RAM
- Integrated Intel HD Graphics 520 (no CUDA support → training was CPU-bound and slower)
-
Limitations Encountered:
- CPU-only training significantly increased runtime.
- Limited RAM availability (16 GB) constrained batch sizes and prolonged convergence.
- No dedicated GPU → unsuitable for scaling to larger datasets or higher-resolution GANs.
-
Alternative Setup (recommended for faster training):
Ubuntu 12.10 — Amazon EC2 Instance — High-Memory Quadruple Extra Large (m2.4xlarge).
This environment offers sufficient memory and GPU acceleration for practical GAN training.