WGAN-GP

PyTorch implementation of WGAN with gradient penalty introduced in the paper: WGAN, WGAN-GP

Hyperparameters

All hyperparameters of this implementation specified in config file config.py

Dataset

Original Dataset: https://www.kaggle.com/soumikrakshit/anime-faces

Results

Training time 1h 30m 48s (GPU Tesla P100-PCIE-16GB) (10 epochs)

Interpolation

ARGS and runs

optional arguments:
  --data_path            path to dataset folder
  --seed                 seed value, default=7889
  --checkpoint_path      path to checkpoint.pth.tar
  --out_path             path to output folder
  --resume_id            wandb id of project for resume metric
  --device               use device, can be - cpu, cuda, tpu, if not specified: use gpu if available

  Other paths and other parameters you can set up in config.py

for example: python3 train.py --data_path 'anime_dataset'

Inference

optional arguments:
  --path_ckpt            Path to checkpoint of model
  --num_samples          Number of samples
  --steps                Number of step interpolation
  --device               cpu or gpu
  --out_path             Path to output folder, default=save to project folder
  --gif                  reate gif
  --grid                 Draw grid of images
  --z_size               The size of latent space, default=128
  --img_size             Size of output image
  --resize               if you want to resize images