Skip to content

Latest commit

 

History

History
93 lines (62 loc) · 5.46 KB

README.md

File metadata and controls

93 lines (62 loc) · 5.46 KB

StackGAN

Code for reproducing main results in the paper StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks by Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaolei Huang, Xiaogang Wang, Dimitris Metaxas.

Dependencies

python 2.7

TensorFlow 0.11

[Optional] Torch is needed, if use the pre-trained char-CNN-RNN text encoder.

[Optional] skip-thought is needed, if use the skip-thought text encoder.

In addition, please add the project folder to PYTHONPATH and pip install the following packages:

  • prettytensor
  • progressbar
  • python-dateutil
  • easydict
  • pandas
  • torchfile

Data

  1. Download our preprocessed char-CNN-RNN text embeddings for birds and flowers and save them to Data/.
  • [Optional] Follow the instructions reedscot/icml2016 to download the pretrained char-CNN-RNN text encoders and extract text embeddings.
  1. Download the birds and flowers image data. Extract them to Data/birds/ and Data/flowers/, respectively.
  2. Preprocess images.
  • For birds: python misc/preprocess_birds.py
  • For flowers: python misc/preprocess_flowers.py

Training

  • The steps to train a StackGAN model on the CUB dataset using our preprocessed data for birds.
    • Step 1: train Stage-I GAN (e.g., for 600 epochs) python stageI/run_exp.py --cfg stageI/cfg/birds.yml --gpu 0
    • Step 2: train Stage-II GAN (e.g., for another 600 epochs) python stageII/run_exp.py --cfg stageII/cfg/birds.yml --gpu 1
  • Change birds.yml to flowers.yml to train a StackGAN model on Oxford-102 dataset using our preprocessed data for flowers.
  • *.yml files are example configuration files for training/testing our models.
  • If you want to try your own datasets, here are some good tips about how to train GAN. Also, we encourage to try different hyper-parameters and architectures, especially for more complex datasets.

Pretrained Model

  • StackGAN for birds trained from char-CNN-RNN text embeddings. Download and save it to models/.
  • StackGAN for flowers trained from char-CNN-RNN text embeddings. Download and save it to models/.
  • StackGAN for birds trained from skip-thought text embeddings. Download and save it to models/ (Just used the same setting as the char-CNN-RNN. We assume better results can be achieved by playing with the hyper-parameters).

Run Demos

  • Run sh demo/flowers_demo.sh to generate flower samples from sentences. The results will be saved to Data/flowers/example_captions/. (Need to download the char-CNN-RNN text encoder for flowers to models/text_encoder/. Note: this text encoder is provided by reedscot/icml2016).
  • Run sh demo/birds_demo.sh to generate bird samples from sentences. The results will be saved to Data/birds/example_captions/.(Need to download the char-CNN-RNN text encoder for birds to models/text_encoder/. Note: this text encoder is provided by reedscot/icml2016).
  • Run python demo/birds_skip_thought_demo.py --cfg demo/cfg/birds-skip-thought-demo.yml --gpu 2 to generate bird samples from sentences. The results will be saved to Data/birds/example_captions-skip-thought/. (Need to download vocabulary for skip-thought vectors to Data/skipthoughts/).

Examples for birds (char-CNN-RNN embeddings), more on youtube:

Examples for flowers (char-CNN-RNN embeddings), more on youtube:

Save your favorite pictures generated by our models since the randomness from noise z and conditioning augmentation makes them creative enough to generate objects with different poses and viewpoints from the same discription 😃

Citing StackGAN

If you find StackGAN useful in your research, please consider citing:

@article{han2016stackgan,
  title={StackGAN: Text to Photo-realistic Image Synthesis with Stacked Generative Adversarial Networks},
  author={Han Zhang and Tao Xu and Hongsheng Li and Shaoting Zhang and Xiaolei Huang and Xiaogang Wang and Dimitris Metaxas},
  journal={arXiv:1612.03242},
  year={2016}
}

References

  • Generative Adversarial Text-to-Image Synthesis Paper Code
  • Learning Deep Representations of Fine-grained Visual Descriptions Paper Code