Recurrent-Affine-Transformation-for-Text-to-image-Synthesis

Official Pytorch implementation for our paper Recurrent-Affine-Transformation-for-Text-to-image-Synthesis

Examples

---

Requirements

• python 3.8
• Pytorch 1.11.0+cu113
• easydict
• nltk
• scikit-image
• A 2080 TI (set nf=32 in *.yaml) or a 3090 32GB (set nf=64 in *.yaml)

Note that nf=32 produces a IS around 5.0 on CUB. To reproduce the final results， please use a GPU more than 32GB.

Installation

Clone this repo.

git clone https://github.com/senmaoy/RAT-GAN.git
cd RAT-GAN/code/

Datasets Preparation

1. Download the preprocessed metadata for birds coco and save them to data/
2. Download the birds image data. Extract them to data/birds/.Raw text data of CUB dataset is avaiable here
3. Download coco dataset and extract the images to data/coco/
4. Download flower dataset and extract the images to data/flower/. Raw text data of flower dataset is avaiable here

Note that flower dataset is a bit different from cub and coco with a standalone dataset processing script.

It's easy to train on your own Datasets (similar to the processing for flower dataset)

1. Prepare a captions.pickle containing all the image paths. Note that captions.pickle should be prepared by yourself.
2. Save captions.pickle under data_dir.
3. Put all the captions of an image in a standalone txt file (one caption per line). This txt file will be later read by dataset_flower.py in line 149: cap_path = '%s/%s.txt' % ('/home/yesenmao/dataset/flower/jpg_text/', filenames['img'][i])
4. Run main.py as usual. Dataset_flower.py will automatically process your own dataset.

Pre-trained text encoder

1. Download the pre-trained text encoder for CUB and save it to ../bird/
2. Download the pre-trained text encoder for coco and save it to ../bird/
3. Download the pre-trained text encoder for flower and save it to ../bird/

---

Training

Train RAT-GAN models:

• For bird dataset: python main.py --cfg cfg/bird.yml

• For coco dataset: python main.py --cfg cfg/coco.yml

• For flower dataset: python main.py --cfg cfg/flower.yml

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

Evaluating

Dwonload Pretrained Model

• RAT-GAN for bird. Download and save it to models/bird/
• RAT-GAN for coco. Download and save it to models/coco/
• RAT-GAN for flower. Download and save it to models/flower/

Evaluate RAT-GAN models:

• To evaluate our RAT-GAN on CUB, change B_VALIDATION to True in the bird.yml. and then run python main.py --cfg cfg/bird.yml
• To evaluate our RAT-GAN on coco, change B_VALIDATION to True in the coco.yml. and then run python main.py --cfg cfg/coco.yml
• We compute inception score for models trained on birds using StackGAN-inception-model.
• We compute FID for CUB and coco using (https://github.com/senmaoy/Inception-Score-FID-on-CUB-and-OXford.git).

---

Citing RAT-GAN

If you find RAT-GAN useful in your research, please consider citing our paper:

@article{ye2022recurrent,
  title={Recurrent Affine Transformation for Text-to-image Synthesis},
  author={Ye, Senmao and Liu, Fei and Tan, Minkui},
  journal={arXiv preprint arXiv:2204.10482},
  year={2022}
}

If you are interseted, join us on Wechat group where a dozen of t2i partners are waiting for you! If the QR code is expired, you can add this wechat: Unsupervised2020

The code is released for academic research use only. Please contact me through [email protected]

Reference

• DF-GAN: DF-GAN: A Simple and Effective Baseline for Text-to-Image Synthesis [code]
• StackGAN++: Realistic Image Synthesis with Stacked Generative Adversarial Networks [code]
• AttnGAN: Fine-Grained Text to Image Generation with Attentional Generative Adversarial Networks [code]
• DM-GAN: Realistic Image Synthesis with Stacked Generative Adversarial Networks [code]