Generative Adversarial Network

In this project, I experimented with getting a Generative Adversarial Network (GAN) to work.

---

📌 Project Description

Setup: I used PyTorch and PyTorch Lightning as the framework. The datasets used were:

• MNIST (handwritten digit dataset)
• IRIS Flower dataset
• A custom GHIBLI dataset (hand-drawn landscape images from Studio Ghibli)

Goal: The objective was to generate images that resemble the dataset inputs. For instance, when training on the MNIST dataset, the aim was to generate images that look like handwritten digits.

Model Architectures: I explored different architectures and fine-tuned them. The models used included:

• Generative Adversarial Network (GAN): A basic architecture using a Multi-Layer Perceptron (MLP), with a Discriminator that attempts to detect fake images and a Generator that attempts to produce realistic-looking images to fool the Discriminator.

• Deep Convolutional Generative Adversarial Network (DCGAN): Uses convolutional layers (commonly used in image classification) to improve image realism.

• Conditional Deep Convolutional GAN (CDC-GAN): Uses both the image and metadata (e.g., labels) to condition the generation process.

Training Tricks & Techniques:

• Instance Noise

• Soft labels

• Experience replay

• More tricks:

  • Soumith’s GAN hacks
  • GAN training pitfalls and tips
  • Training GANs faster & better

More Resources:

• Original GAN paper (Goodfellow)
• DCGAN paper
• Conditional GAN paper
• PyTorch DCGAN tutorial

---

▶️ Usage

1. Download the Jupyter Notebook files (.ipynb) you want to use from the code folder.
2. Run them with Jupyter Notebook/Lab or Google Colab.

Notes:

• On Google Colab, all necessary packages should auto-install.
• On a local Jupyter Notebook, you’ll need to install required packages manually.
• Please read the Disclaimer below.

---

⚠️ Disclaimer

This project was created a while ago and the documentation is not very organized. As a result, there are multiple Jupyter Notebooks with vague naming, and differences are not always clearly explained.

When selecting a notebook, inspect both the filename and the code.

Some notebooks may require tinkering before they work. Here’s why:

• Some notebooks do not train from scratch (Epoch 0) but attempt to load from a checkpoint (e.g., Epoch 50 or 100). → To fix: comment out resume_from_checkpoint="" in the pl.Trainer() call.

• Some models are trained using a custom GHIBLI dataset. → To use another dataset, modify the train_dataloader and test_dataloader functions. Be mindful of image sizes when doing so.

• The logger used is comet_ml, but I removed my API key. → Replace with your own comet_ml API key or use a different logging method.

• There may be other issues I don’t currently remember.

---

✅ Results

I was able to generate convincing digit images in various resolutions, including 16×16, 32×32, 64×64, and 128×128 (via upscaled MNIST). However, results for the flower and GHIBLI datasets were less successful.

✔ Implemented Models

• Standard MLP (GAN on MNIST)

• Standard DCGAN

• DCGAN with training tricks

• DCGAN with colored GHIBLI landscape images (64×64)

• DCGAN with flower images (128×128)

📉 Loss Graphs

• Discriminator Loss

• Generator Loss

👉 Click here to see more results and loss graphs

---

📄 License

Distributed under the MIT License. See LICENSE for more information.