Read the paper here.
- Clone this repository
- Install pip packages:
- torch
- torchvision
- tensorflow (for FID/IS calculation)
- wandb
- numpy
- scipy
- tqdm
- imageio
- For Weights and Biases (wandb), follow install/setup instructions here. If you want to disable wandb, you can set the env variable
WANDB_MODE=dryrun
, although there is no alternate logging. - Download the FID statistics of the test set following the instructions here. Downlaod to
[DATA_ROOT]/fid_stat/fid_stats_[DATASET]_train.npz
, whereDATA_ROOT
defaults to.
.
Train a model by running a command a command similar to the following from the root directory of this repository:
python -m gr_gan.launch --dataset=cifar10 --model=resnet --loss=nonsaturating --lr=.0001 --lr_dis=.0004 --z_dim=128 --batch_size=64 --train_fn=baseline --evaluate_freq=5000 --iterations=100000 --log_freq=10 --spectral_norm=True
Descriptions of all command-line options are available by running:
python -m gr_gan.trainable --help