Text-Adaptive Generative Adversarial Network (TAGAN)

A PyTorch implementation of the paper "Text-Adaptive Generative Adversarial Networks: Manipulating Images with Natural Language". This code implements a Text-Adaptive Generative Adversarial Network (TAGAN) for manipulating images with natural language.

Requirements

Pretrained word vectors for fastText

Download a pretrained English word vectors. You can see the list of pretrained vectors on this page.

Datasets

Oxford-102 flowers: images and captions
Caltech-200 birds: images and captions

The caption data is from this repository. After downloading, modify CONFIG file so that all paths of the datasets point to the data you downloaded.

Pretrained models

Oxford-102 flowers: flowers_G.pth
Caltech-200 birds: birds_G.pth

Please put these files in ./models/ folder.

Run

scripts/preprocess_caption.sh
Preprocess caption data using fastText embedding. You only need to run it once before training.
scripts/train_[flowers/birds].sh
Train a network. If you want to change arguments, please refer to train.py.
scripts/test_[flowers/birds].sh
Test a trained network. After running it, please see ./test/result_[flowers/birds]/index.html.

Results

Citation

Please cite our paper when you use this code.

@inproceedings{nam2018tagan,
  title={Text-Adaptive Generative Adversarial Networks: Manipulating Images with Natural Language},
  author={Nam, Seonghyeon and Kim, Yunji and Kim, Seon Joo},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
  year={2018}
}

Contact

Please contact shnnam@yonsei.ac.kr if you have any question about this work.

MapleSpirit/tagan