improved_LSGAN

Code for On the Effectiveness of Least Squares Generative Adversarial Networks and Least Squares Generative Adversarial Networks

Related project: LSGAN

Prerequisites

• Tensorflow
• SciPy

Cats

The code of pre-process methods is from this project.

1. Prepare data

cd cats
sh convert_data.sh

2. Training

python main.py --loss lsgan
python main.py --loss dcgan

3. Sampling from saved models

   Download saved models from here.

cat saved_models.tar.gz.* | tar xzvf -
python sampling --checkpoint_dir ./saved_models/lsgan/
python sampling --checkpoint_dir ./saved_models/dcgan/

Datasets with small variance

0. If using the dataset only

cd small_variance_datasets
tar xzvf data.tar.gz
#Then find the dataset in ./data/

1. Prepare data

cd small_variance_datasets
sh convert_data.sh

2. Training

python main --loss lsgan
python main --loss dcgan

Difficult architectures

The code of this experiment is based on this project.

1. Prepare data

   Download LSUN-bedroom

cd difficult_architectures/resnet
sh convert_data.sh $DATA_DIR

2. Training

python gan_64x64.py

Citation

If you use this work in your research, please cite:

@article{arxiv1712.06391,
  author = {Xudong Mao and Qing Li and Haoran Xie and Raymond Y.K. Lau and Zhen Wang and Stephen Paul Smolley},
  title = {On the Effectiveness of Least Squares Generative Adversarial Networks},
  journal = {arXiv preprint arXiv:1712.06391},
  year = {2017}
}
@inproceedings{Mao2017,
  author = {Xudong Mao and Qing Li and Haoran Xie and Raymond Y.K. Lau and Zhen Wang and Stephen Paul Smolley},
  title = {Least Squares Generative Adversarial Networks},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year = 2017
}