IntroVAE-Pytorch

Pytorch Implementation for NeuraIPS2018 paper: IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis.

The rep. contains a basic implementation for IntroVAE. However, due to no official implementation released, some hyperparameters can only be guessed and can not reach the performance as stated in paper.

HowTo

1. Download CelebA dataset and extract it as:

├── /home/i/dbs/
	├──img_align_celeba # only one folder in this directory
		├── 050939.jpg
		├── 050940.jpg
		├── 050941.jpg
		├── 050942.jpg
		├── 050943.jpg
		├── 050944.jpg
		├── 050945.jpg

modify /home/i/dbs to your specific path, making sure that the /home/i/dbs/ comtains only ONE folder since we use torchvision.datasets.ImageFolder API to load dataset.

    argparser.add_argument('--root', type=str, default='/home/i/dbs/',
                           help='root/label/*.jpg')

2. run python main.py --epoch 750000 to train from strach, and use python main.py --resume '' --epoch 1000000 to resume training from latest checkpoint.

Training

only tested for CelebA 128x128 exp.

• training curves

• sampled x