Project Overview
This project marked my introduction to deep learning. As my very first deep learning project, I chose to explore the world of Generative Adversarial Networks (GANs) by training models to generate realistic cat images. I was accompanied by a more experienced member of Automatants, the AI student association of CentraleSupélec. The project had a simple goal: train a model to generate images of cats.
Generated Results
The journey from concept to working model was filled with both breakthroughs and setbacks. Here are some examples of the generated cat images produced by the different model architectures:
Technical Approach
1. Deep Convolutional GAN (DCGAN)
I began with a Deep Convolutional Generative Adversarial Network (DCGAN). I simply took a tutorial on GANs from TensorFlow. This relatively simple model used convolutional layers in both the generator and discriminator, providing a working starting point for image generation.
2. Enhanced Architecture with Residual Connections
To improve model performance, I experimented with adding residual connections to the GAN architecture. I had in mind that larger and deeper networks would be able to generate better images.
3. Progressive Growing GAN
Then, I wanted to dive into more complexity. Since StyleGAN seemed a bit too complex, I decided to implement a Progressive GAN. The idea is to starts training on very small images (4×4 pixels) and gradually increases the resolution during training. The progressive nature allows the model to learn coarse features first, then refine them to finer details, theoretically producing more stable training and higher quality results. In practice, the implementation worked but I did not see any significant gain in quality.
Implementation and Deployment
Dataset and Preprocessing
The project utilized a curated dataset of cat images sourced from the internet. The images underwent standard preprocessing steps including resizing, normalization, and augmentation to ensure consistent input for the neural networks.
Public Deployment
To share the project with a broader audience, I deployed the trained models using TensorFlow.js, enabling real-time cat generation directly in web browsers. This interactive deployment allows users to experience the generative capabilities firsthand.
Try the live demo: Cat Generator
Reflection and thoughts
I was extremely excited when I was able to generate cat images. I loved the idea of being able to produce something by myself. I learned a lot about programming and deep learning. And even more importantly, I started to learn from blogs, papers, Github repositories, etc. Now (in 2023), I should have done so many things differently. The results are not very good and I do not even know if my models converged properly.
