xuezhongcailian / TIC

[DCC 2022] Transformer-based Image Compression

Transformer-based Image Compression

Pytorch Implementation of "Transformer-based Image Compression"[arXiv], DCC 2022.

Our newly released work "TinyLIC" with more efficient performance can be found at the homepage.

Acknowledgement

The framework is based on CompressAI, we add our networks in compressai.models.tic and compressai.layers for usage.

Installation

To get started locally and install the development version of our work, run the following commands (The docker environment is recommended):

git clone https://github.com/lumingzzz/TIC.git
cd TIC
pip install -U pip && pip install -e .
pip install timm

Usage

Train

We use the Flicker2W dataset for training, and the script for preprocessing.

Run the script for a simple training pipeline:

python examples/train.py -m tic -d /path/to/my/image/dataset/ --epochs 300 -lr 1e-4 --batch-size 8 --cuda --save

Evaluation

An example to evaluate model:

python -m compressai.utils.eval_model checkpoint path/to/eval/data/ -a tic -p path/to/pretrained/model --cuda

Notes

Some implementations are slightly different from the paper:

1. We remove the activation functions after the convolutions (e.g. the GDN and LReLU), which have no influence to the performance.
2. The implementation of the Causal Attention Module (CAM) is slightly different from the paper by directly masking the input of context model, it shows more feasible than the original one.

Citation

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

@INPROCEEDINGS{9810760,
               author={Lu, Ming and Guo, Peiyao and Shi, Huiqing and Cao, Chuntong and Ma, Zhan},
               booktitle={2022 Data Compression Conference (DCC)}, 
               title={Transformer-based Image Compression}, 
               year={2022},
               volume={},
               number={},
               pages={469-469},
               doi={10.1109/DCC52660.2022.00080}}
              }

Contact

If you have any question, please contact me via [e-mail].