github-pengge / PyTorch-progressive_growing_of_gans

PyTorch implementation of Progressive Growing of GANs for Improved Quality, Stability, and Variation.
580 stars 98 forks source link
gan pggan pytorch

PyTorch implementation of Progressive Growing of GANs for Improved Quality, Stability, and Variation.

Warning: the master branch might collapse. To obtain similar result in README, you can fall back to this commit, but remembered that some ops were not correctly implemented under that commit. Besides, you'd better use a lower learning rate, 1e-4 would be fine.

How to create CelebA-HQ dataset

I borrowed h5tool.py from official code. To create CelebA-HQ dataset, we have to download the original CelebA dataset, and the additional deltas files from here. After that, run

python2 h5tool.py create_celeba_hq file_name_to_save /path/to/celeba_dataset/ /path/to/celeba_hq_deltas

This is what I used on my laptop

python2 h5tool.py create_celeba_hq /Users/yuan/Downloads/CelebA-HQ /Users/yuan/Downloads/CelebA/Original\ CelebA/ /Users/yuan/Downloads/CelebA/CelebA-HQ-Deltas

I found that MD5 checking were always failed, so I just commented out the MD5 checking part(LN 568 and LN 589)

With default setting, it took 1 day on my server. You can specific num_threads and num_tasks for accleration.

Training from scratch

You have to create CelebA-HQ dataset first, please follow the instructions above.

To obtain the similar results in samples directory, see train_no_tanh.py or train.py scipt for details(with default options). Both should work well. For example, you could run

conda create -n pytorch_p36 python=3.6 h5py matplotlib
source activate pytorch_p36
conda install pytorch torchvision -c pytorch
conda install scipy
pip install tensorflow

#0=first gpu, 1=2nd gpu ,2=3rd gpu etc...
python train.py --gpu 0,1,2 --train_kimg 600 --transition_kimg 600 --beta1 0 --beta2 0.99 --gan lsgan --first_resol 4 --target_resol 256 --no_tanh

train_kimg(transition_kimg) means after seeing train_kimg * 1000(transition_kimg * 1000) real images, switching to fade in(stabilize) phase. Currently only support LSGAN and GAN with --no_noise option, since WGAN-GP is unavailable, --drift option does not affect the result. --no_tanh means do not use tanh at generator's output layer.

If you are Python 2 user, You'd better add this to the top of train.py since I use print('something...', file=f) to write experiment settings to file.

from __future__ import print_function

Tensorboard

tensorboard --logdir='./logs'

Update history

Reference implementation