jppgks / stackgan-pp

TensorFlow implementation of StackGAN++, as described in the paper by Zhang, Xu et al.

StackGAN++

TensorFlow implementation of StackGAN++, described in the paper by Zhang, Xu et al.

Run project

Dependencies. Python 3, TensorFlow 1.4 (, TensorBoard)

Data. You can find used datasets on FloydHub:

• Text embeddings
• Bird images

Training.

1. Clone the repo, including the TensorFlow models submodule:

   git clone --recurse-submodules https://github.com/jppgks/stackgan-pp.git

2. Run the training script

   python ./train.py

   optionally with arguments. All possible arguments, with their doc strings, are listed when running:

   python ./train.py --help

3. Follow progress in TensorBoard:

   tensorboard --logdir=<TRAIN_LOG_DIR location>

Project structure

The project aims to reproduce StackGAN++ paper results by introducing as little modifications as possible to the existing TFGAN framework.

TFGAN → TFSTACKGAN. tfstackgan mimics the folder structure of TFGAN. tfstackgan/python/train.py contains modified TFGAN train.py functions. The color loss for the generator is defined in tfstackgan/python/losses/python/losses_impl.py. The ./train.py and ./networks.py scripts are modeled after the TFGAN CIFAR example.

Comparison with paper

Upsampling. This implementation does not make use of GLUs and/or residual blocks at the moment. Upsampling in all generator stages happens through fractionally strided convolutions.

References

• TFGAN
• StackGAN++ PyTorch implementation