High-Fidelity Variable-Rate Image Compression via Invertible Activation Transformation

Installation

We tested our code in Ubuntu 18.04.5 LTS, g++ 7.5.0, cuda 11.1, python 3.8.11, pytorch 1.8.2.

Install needed packages.

• $ pip install torch==1.8.2+cu111 torchvision==0.9.2+cu111 torchaudio==0.8.2 -f https://download.pytorch.org/whl/lts/1.8/torch_lts.html
• $ pip install -r requirements.txt
• $ pip install compressai==1.1.8

Dataset

• Training set: Flicker 2W about 20745 images or here. Run the data preprocessing file ./scripts/flicker_process.py
• Test set: Kodak dataset, CLIC, DIV2K

Run following command in scripts directory to create your_trainset.csv and your_testset.csv.

• $ python get_paths.py --dir=your_train_dataset_dir/ --name=your_trainset
• $ python get_paths.py --dir=your_test_dataset_dir/ --name=your_testset
• your_trainset.csv and your_testset.csv will be created in data directory.

Training

Configuration

We used the configuration as ./configs/config.yaml to train our model.

Train

$ python train.py --config=./configs/config.yaml --name=your_instance_name
The checkpoints of the model will be saved in ./results/your_instance_name/snapshots.
Training for 2.5M iterations will take about 3 weeks on a single GPU like NVIDIA RTX3090. At least 12GB GPU memory is needed for the default training setting.

Evaluation

$ python eval.py --snapshot=./results/your_instance_name/snapshots/best.pt --testset=./data/your_testset.csv We release the pretrained model. The extracted password is: HFVR. You can download and put it in results directory.

Note that as mentioned in original CompressAI, "Inference on GPU is not recommended for the autoregressive models (the entropy coder is run sequentially on CPU)." So for inference of our model, please run on CPU.

Citation

If you find this work useful for your research, please cite:

@inproceedings{cai2022high,
  title={High-Fidelity Variable-Rate Image Compression via Invertible Activation Transformation},
  author={Cai, Shilv and Zhang, Zhijun and Chen, Liqun and Yan, Luxin and Zhong, Sheng and Zou, Xu},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
  pages={2021--2031},
  year={2022}
}

@article{cai2024I2C,
  author={Cai, Shilv and Chen, Liqun and Zhang, Zhijun and Zhao, Xiangyun and Zhou, Jiahuan and Peng, Yuxin and Yan, Luxin and Zhong, Sheng and Zou, Xu},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={I2C: Invertible Continuous Codec for High-Fidelity Variable-Rate Image Compression}, 
  year={2024},
  volume={},
  number={},
  pages={1-18}