ZJU_summer_dpl_23_2

DifGAN PyTorch Implementation

This is the repository of the course project implement of a summer course called "Cutting-edge Technologies and Applications of Deep Learning" taught by Hehe Fan in ZJU. We gave four methods in ./train, they are original GAN, high dimension GAN, reimplementation of diffusion GAN in Diffusion-GAN: Training GANs with Diffusion, and our own DifGAN respectively.

./models includes the models of discriminator and generator (with and without tilmestep as input).

./result should record the generated images and the model parameters(.pth) after training.

There is no ./dataset folder as we use Mnist in official library.

Train DifGAN

• 64-bit Python 3.7 or newer version
• PyTorch 1.7.1 or newer version. See https://pytorch.org/ for PyTorch install instructions.
• CUDA or no CUDA are both available (only CPU would be slow though)

Data Preparation

In the project we trained our model on CIFAR (32 x 32) and MNIST (28 x 28) It is not necessary to download the dataset separately, the downloading is embedded in current code

Training

e.g. If you want to train gan model, you can run python -m train.gan under project directory