GAN_generate

generate GAN pics based on StudioGAN

I have push the ckpt and other "BIG" files using git Large File Storage (FLS), you may need to pull it using lfs too.

requirement

First, install PyTorch meeting your environment (at least 1.7):

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu116

Then, use the following command to install the rest of the libraries:

pip install tqdm ninja h5py kornia matplotlib pandas sklearn scipy seaborn wandb PyYaml click requests pyspng imageio-ffmpeg timm

Run

Get the free gpu cuda number

nvidia-smi

then

./generate [CUDA] [MODEL] [number of image to save] [version(optional)]

for example

./generate 0 BigGAN-Deep 100

will generate BigGAN-Deep models with random ckpt version running on CUDA 0, and save 100 generated pics

./generate 0 BigGAN-Deep 10 56_49

will generate BigGAN-Deep models with the first ckpt version having the name "56_49" running on CUDA 0, and save 10 pics

The generated pics will locate on ./samples/MODEL_FOLDER/fake/*/*.png

If you need anything, contact me directly which will help me making this readme more readable and friendly.