daysonn / efficient-t2i

Official implementation of the paper Efficient Neural Architecture for Text-to-Image Synthesis.

Efficient T2I

Official implementation of the paper Efficient Neural Architecture for Text-to-Image Synthesis.

Requirements

• Python 3.7+
• Pytorch 1.2+
• Tensorflow 1.14 (pip install tensorflow-gpu==1.14) (used to compute IS and FID)
• gdown (pip install gdown) (used to download datasets and meta data from Google Drive)
• easydict (pip install easydict)
• tensorboardX (pip install tensorboardx)
• tqdm (pip install tqdm)

Datasets

To download CUB: ./scripts/download_birds.sh

To download Oxford-102: ./scripts/download_flowers.sh

Training

To train:

./scripts/train_birds.sh

Please look at the script for setting training parameters.

After launching a training job, follow it on tensorboard. Go to the project folder then:

tensorboard --logdir=logs/

Evaluation

To eval:

./scripts/eval_cub.sh

Please look at the script for setting evaluation parameters.

Pretrained Models

We already uploaded the pretrained model for Birds, download it using the provided script:

./scripts/download_pretrained_birds_model.sh

Playground

We also include a jupyter notebook with examples on how to generate images. Just go to the project folder and launch:

$jupyter notebook

Citation

If you find this work useful, please consider citing:

@article{souza2020efficient,
  title={Efficient Neural Architecture for Text-to-Image Synthesis},
  author={Souza, Douglas M and Wehrmann, J{\^o}natas and Ruiz, Duncan D},
  journal={arXiv preprint arXiv:2004.11437},
  year={2020}
}