AttnGAN-CelebA

Pytorch implementation for AttnGAN with CelebA dataset.

Architecture

Dependencies

• python 2.7
• Pytorch
• In addition, please add the project folder to PYTHONPATH and pip install the following packages:
  • python-dateutil
  • easydict
  • pandas
  • torchfile
  • nltk
  • scikit-image
  • pyyaml

Data

• Download our preprocessed text for CelebA and extract them to data/CelebA/
  • file directory example: data/CelebA/text/0/000012.txt
• Download the preprocessed CelebA image data and extract them to data/CelebA/
  • file directory example: data/CelebA/images/000012.jpg

Training

1. Pre-train DAMSM models:

   For CelebA dataset: python pretrain_DAMSM.py --cfg cfg/DAMSM/CelebA.yml --gpu 0

2. Train AttnGAN models:

   For CelebA dataset: python main.py --cfg cfg/CelebA_attn2.yml --gpu 0

• *.yml files are example configuration files for training/evaluation our models.

Sampling

• Run python main.py --cfg cfg/eval_CelebA.yml --gpu 0 to generate examples from captions in files listed in "./data/CelebA/example_filenames.txt". Results are saved to DAMSMencoders/.
• Input your own sentence in "./data/CelebA/example_captions.txt" if you want to generate images from customized sentences.

Validation

• To generate images for all captions in the validation dataset, change B_VALIDATION to True in the eval_*.yml. and then run python main.py --cfg cfg/eval_CelebA.yml --gpu 0

result - attention map

• 60epoch

Reference

• AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks