README.md

Chainer-StackGAN

Chainer implementation of StackGAN on CUB

"StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks," Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaolei Huang, Xiaogang Wang, Dimitris Metaxas, arXiv:1612.03242. https://arxiv.org/abs/1612.03242
Original code: https://github.com/hanzhanggit/StackGAN

Requirements

• Python 2.7.6+
  • Chainer 1.22.0+
  • numpy 1.12.1+
  • scipy 0.13.3+

Test

Quick Demo

1. Download text features from https://drive.google.com/open?id=0B3y_msrWZaXLT1BZdVdycDY5TEE and save it to data/
2. Download pretrained weights from https://drive.google.com/drive/folders/0B6nc1VY-iOPZU0w2OVZBMGx3cTQ?usp=sharing and save them to models/
3. cd stage2
4. python demo.py -g=0

Pretrained Weights

Download pretrained weights from my google drive https://drive.google.com/drive/folders/0B6nc1VY-iOPZU0w2OVZBMGx3cTQ?usp=sharing Put them on models/

Example result on CUB

• stage1
  
• stage2
  

todo

• multi gpu

Dataset

See original code of StackGAN to prepare CUB dataset

python StackGAN/misc/preprocess_birds.py

Train

cd stage1
python train_stackgan.py -g=0 
cd stage2
python train_stackgan.py -g=0 -s1w=path/to/stage1/gen_epoch_600.npz