Continual Cross-domain Image Compression via Entropy Prior Guided Knowledge Distillation and Scalable Decoding

[Uncompleted] Pytorch code for the paper "Continual Cross-domain Image Compression via Entropy Prior Guided Knowledge Distillation and Scalable Decoding".

TODO:

• Datasets
• Models
• Train code
• Evaluation code

This repository is based on CompressAI.

Installation

conda create -n ccic python=3.7
conda activate ccic
pip install compressai
pip install numpy==1.19.1
pip install tqdm

Usage

Training

CUDA_VISIBLE_DEVICES='1' python train.py --model mbt2018 --filters 128 192 --lambda 0.032 --checkpoints_dir /path/to/checkpoint/folder/ --epochs 10000 10000 10000 10000 10000 --save_interval 200 -lr 1e-4

Datasets

• CLIC [https://www.compression.cc/]
• OP [Unsplash Dataset | The world’s largest open library dataset]
• GTA5 [https://download.visinf.tu-darmstadt.de/data/from_games/]
• DOTA [https://captain-whu.github.io/DOTA/index.html]
• COVID19 [https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=89096912#89096912aaf792e54ffd4c16ad8dd5b4b117ab3f]