A pytorch implementation of Nvidias StyleGan for the class of Deep Unsupervised Learning at LMU Munich 2021
To run training execute the file dataset.py first. For example:
!python dataset.py --dataset_path --out_path --max_size --min_size --magic_number
dataset_path corresponds to the file path where the dataset is located.
Has to be of the structure of a torchvision imagefolder:
https://pytorch.org/vision/stable/datasets.html#torchvision.datasets.ImageFolder
out_path is the file path of the lmdb database that will be generatedd
max_size is the maximum image size (default 128x128)
min_size is the minimum image size (default 8x8)
magic_number is a multiplying constant to initialize the data base in the correct size on windows machines.
After preprocessing the images StyleGAN can be trained with:
!python train.py path --init_size --max_size --mixing --loss
Where path is the path to the lmdb database created above
init_size is the size at which to start progressively growing (default 8x8)
max_size is the maximum size of the training images (default 128x128)
mixing specifies that mixing regularization will be used
loss specifies the gan loss used (r1 or wgan-gp)
to train stylegan2-ada the official pytorch implementation was used,
to evaluate the checkpoints for interpolation only the model2 files are necessary to be loaded like they are
in the respective notebooks.