This is a simplified PyTorch implementation of StyleGAN2 introduced in Analyzing and Improving the Image Quality of StyleGAN paper
The main purpose is to implement a smaller version of the StyleGAN using another interpolation method and evaluate the quality of result model built by my own.
This is a study project to understand what is GANs, how they can be designed and how can be trained.
The code is based on https://github.com/labmlai/annotated_deep_learning_paper_implementations/tree/master/labml_nn/gan/stylegan
- PyTorch
- Numpy
- scikit-learn
- matplotlib (just to visualize training results)
- Santiel 5K Truecolor Landshapes-4041
- Because of small computational power the model can't be trained longer to achieve desired image quality
- Currently looking for an opportunity to run model in distributed mode (multiple gpu's/multiple nodes/combined)
- A simple web-based API for image generation is currently unimplemented.
- The network generates 32
$\times$ 32 px instead of 1024$\times$ 1024 px. - Smaller architecture: mapping network (5 layers used instead of 8 in original), generator and discriminator (
$\log_2(32) = 5$ blocks instead of$\log_2(1024) = 10$ ) - Noise injection was performed in two steps:
-
- Generate the base noise before mapping styles.
-
- In each style block add randomly generated noise instead of using pregenerated outter.
-
- Default interpolation method is
nearest-exactinstead ofbilinearso produced images are more smooth.
 |
|---|
| Image 1 - Generator and discriminator train losses per epochs |
 |
|---|
| Image 2 - training results per epochs: 1200 epochs (a), 1500(b), 3200(c), 3500(c), 4200(e) |
 |
|---|
| Image 3 - Comparison of zoomed generated (left) and real (right) images |