Simple Pytorch implementation of Stylegan2 that can be completely trained from the command-line, no coding needed. You will need a machine with a GPU and CUDA installed. You can also specify the location where intermediate results and model checkpoints should be stored. You can increase the network capacity (which defaults to 16) to improve generation results, at the cost of more memory. By default, if the training gets cut off, it will automatically resume from the last checkpointed file. Once you have finished training, you can generate images from your latest checkpoint. If a previous checkpoint contained a better generator, (which often happens as generators start degrading towards the end of training), you can load from a previous checkpoint with another flag. A technique used in both StyleGAN and BigGAN is truncating the latent values so that their values fall close to the mean. The small the truncation value, the better the samples will appear at the cost of sample variety.
Features
- Multi-GPU training
- Low amounts of Training Data
- This framework also allows for you to add an efficient form of self-attention to the designated layers of the discriminator
- The more GPU memory you have, the bigger and better the image generation will be
- Nvidia recommended having up to 16GB for training 1024x1024 images
- Deployment on AWS
Project Samples
