Comprehensive Comparisons of Uniform Quantizers for Deep Image Compression

Official implementation of "Comprehensive Comparisons of Uniform Quantizers for Deep Image Compression" in ICIP 2021.

Environment

• CUDA==10.0
• CUDNN==7.6.0
• Python

pip install pipenv
pipenv install

Refer to Pipfile if you download packages manually.

Usage

# train a model
python main.py --verbose --checkpoint_dir checkpoints/l0.01_aun_aun --qua_ent AUN-Q train --lambda 0.01 --qua_dec AUN-Q --train_root /path/to/ImageNet/train/

# evaluate the model
python evaluate.py /path/to/Kodak/images/ --qua_ent AUN-Q --checkpoint_dir checkpoints/l0.01_aun_aun/

You can train other combinations of approximation methods by specifying --qua_ent and --qua_dec.

Please select from {AUN-Q, STE-Q, U-Q, SGA-Q} for each option.

Citation

@inproceedings{tsubotaICIP21,
    title = {Comprehensive Comparisons of Uniform Quantizers for Deep Image Compression},
    author = {Tsubota, Koki and Aizawa, Kiyoharu},
    booktitle = {ICIP},
    year = {2021},
    pages={2089-2093}
}

Note

Our code is based on the example codes in Tensorflow Compression licensed under Apache-2.0 (Copyright 2018 Google LLC).