Translation-equivariant Image Quantizer for Bi-directional Image-Text Generation

Woncheol Shin¹, Gyubok Lee¹, Jiyoung Lee¹, Joonseok Lee^2,3, Edward Choi¹ | Paper

¹KAIST, ²Google Research, ³Seoul National University

Abstract

Recently, vector-quantized image modeling has demonstrated impressive performance on generation tasks such as text-to-image generation. However, we discover that the current image quantizers do not satisfy translation equivariance in the quantized space due to aliasing, degrading performance in the downstream text-to-image generation and image-to-text generation, even in simple experimental setups. Instead of focusing on anti-aliasing, we take a direct approach to encourage translation equivariance in the quantized space. In particular, we explore a desirable property of image quantizers, called 'Translation Equivariance in the Quantized Space' and propose a simple but effective way to achieve translation equivariance by regularizing orthogonality in the codebook embedding vectors. Using this method, we improve accuracy by +22% in text-to-image generation and +26% in image-to-text generation, outperforming the VQGAN.

Requirements

conda env create -f environment.yaml
conda activate bidalle
pip install horovod==0.22.1

If you fail to install horovod, please refer to here.

Download Dataset

bash download_mnist64x64_stage2.sh

Download Image Classifier

bash download_classifier_ckpt.sh

Training Bi-directional Image-Text Generator (Stage 2)

In run_train_dalle.sh, you should specify --vqgan_model_path and --vqgan_config_path. Provide your model path pretrained from TE-VQGAN. For example,

--vqgan_model_path /home/TE-VQGAN/logs/2022-04-01T07-37-39_mnist64x64_vqgan/checkpoints/last.ckpt \
--vqgan_config_path /home/TE-VQGAN/logs/2022-04-01T07-37-39_mnist64x64_vqgan/configs/2022-04-01T07-37-39-project.yaml

And then run the script:

bash run_train_dalle.sh

Citation

@misc{shin2021translationequivariant,
      title={Translation-equivariant Image Quantizer for Bi-directional Image-Text Generation}, 
      author={Woncheol Shin and Gyubok Lee and Jiyoung Lee and Joonseok Lee and Edward Choi},
      year={2021},
      eprint={2112.00384},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgments

The implementation of 'TE-VQGAN' and 'Bi-directional Image-Text Generator' is based on VQGAN and DALLE-pytorch.