pytorch-pretrained-BigGAN

Why we included this project

This is a PyTorch port of DeepMind's BigGAN that tracks the original TensorFlow implementation closely, with output differences on the order of 1e-5. It ships with pretrained generator weights at 128, 256 and 512 pixel resolutions, so you can load a model and produce ImageNet-scale images without training anything yourself. The repo also includes the scripts that pull the weights from TensorFlow Hub and convert them, which saves real time if you have wrestled with TF-to-PyTorch conversions before. One honest caveat: only the generator is included, since DeepMind never released the discriminator weights. For generative art, synthetic training data, or just studying how a large GAN is structured, this is a compact codebase worth reading.