Home

Awesome

QVRF: A QUANTIZATION-ERROR-AWARE VARIABLE RATE FRAMEWORK FOR LEARNED IMAGE COMPRESSION

Official implementation of "QVRF: A QUANTIZATION-ERROR-AWARE VARIABLE RATE FRAMEWORK FOR LEARNED IMAGE COMPRESSION"

Table of Contents

Related links

Environment

Recommend using Miniconda.

    #python>=3.6 should be fine.
    conda create -n qraf python=3.8
    conda activate qraf
    pip install compressai==1.1.5
    #pip install compressai==1.1.5 -i https://pypi.mirrors.ustc.edu.cn/simple/

Dataset

   mkdir dataset
   mv Kodak ./dataset

Download Collection of Kodak or Kodak for Testing.

Training

Using Trainningdataset_Preprocessing.py to select the largest 8000 images from imageNet and 584 images from CLIC2020 and to preprocess the images as the training dataset.

Parameters

training Stage1

python3 train.py -d ./dataset  -e 2000 -lr 1e-4 -n 8 --batch-size 8 --test-batch-size 64 --aux-learning-rate 1e-3 --patch-size 256 256 --cuda --save --seed 1926 --clip_max_norm 1.0  --stage 1 --ste 0  --loadFromPretrainedSinglemodel 0

training Stage2

python3 train.py  -d ./dataset  -e 500 -lr 1e-4 -n 8 --batch-size 8 --test-batch-size 64 --aux-learning-rate 1e-3 --patch-size 256 256 --cuda --save --seed 1926 --clip_max_norm 1.0  --stage 2 --ste 0  --refresh 1 --loadFromPretrainedSinglemodel 0 --checkpoint checkpoint_best_loss.pth.tar |tee Cheng2020Noise.txt

Actually, you can load a fixed-rate model and finetune it with QVRF.

python3 train.py  -d ./dataset  -e 500 -lr 1e-4 -n 8 --batch-size 8 --test-batch-size 64 --aux-learning-rate 1e-3 --patch-size 256 256 --cuda --save --seed 1926 --clip_max_norm 1.0  --stage 2 --ste 0  --refresh 1 --loadFromPretrainedSinglemodel 1 --checkpoint cheng2020_attn-mse-6-730501f2.pth.tar |tee Cheng2020Noise.txt

training Stage3

python3 train.py  -d ./dataset  -e 500 -lr 1e-4 -n 8 --batch-size 8 --test-batch-size 64 --aux-learning-rate 1e-3 --patch-size 256 256 --cuda --save --seed 1926 --clip_max_norm 1.0  --stage 3 --ste 1 --refresh 1 --loadFromPretrainedSinglemodel 0 --checkpoint checkpoint_best_loss.pth.tar |tee Cheng2020STE.txt

Update

Once the model is trained, we need to run update.py to fix the entropy model

python3 update.py checkpoint_best_loss.pth.tar   -n Cheng2020VR

Inference

Download checkpoint variable rate model of Cheng2020 for Inference.

Parameters

Inference code

    python3 Inference.py --dataset TestDataset --s 2 --output_path output_pathName -p CheckpointPath --patch 64 --factormode 1 --factor 0.1

Example of Inference.py

Discrete bitrate results

For all discrete bitrate results:

    python3 Inference.py --dataset ./dataset/Kodak --s 8 --output_path AttentionVRSTE -p ./Cheng2020VR.pth.tar --patch 64 --factormode 0 --factor 0

For discrete bitrate results at a assign Index: Index belongs in {0, 1, 2, 3, 4, 5, 6, 7}

    python3 Inference.py --dataset ./dataset/Kodak --s Index --output_path AttentionVRSTE -p ./Cheng2020VR.pth.tar --patch 64 --factormode 0 --factor 0

Continuous bitrate results

For example continuous bitrate results:

    python3 Inference.py --dataset ./dataset/Kodak --s 2 --output_path AttentionVRSTE -p ./Cheng2020VR.pth.tar --patch 64 --factormode 1 --factor 0.1

Change arbitrary quantization bin size in the range of [0.5, 12] at a quantization bin size 1/QBS. (1/QBS in range of [0.5, 12])

    python3 Inference.py --dataset ./dataset/Kodak --s 2 --output_path AttentionVRSTE -p ./Cheng2020VR.pth.tar --patch 64 --factormode 1 --factor 1/QBS

Note

A higher bitrate corresponds to a larger factor value which is the reciprocal of quantization bin size.

Discrete/Continuous Variable Rate Results

Note

From public code and paper, the models of Cheng2020 only trained for the low and medium rate with lambda belonging to {0.0016, 0.0032, 0.0075, 0.015, 0.03, 0.045}.

We re-trained Cheng2020 on our training dataset following the original paper setting with lambda belonging to {0.0018, 0.0035, 0.0067, 0.0130, 0.0250, 0.0483, 0.0932, 0.1800} for a fair comparison.

Improvement of QVRF

Using the predefined Lagrange multiplier set {0.0018, 0.0035, 0.0067, 0.0130, 0.025, 0.0483, 0.0932, 0.18, 0.36, 0.72, 1.44} with can achieve better results in high bitrate.

You can download the checkpoint for testing the result with the predefined Lagrange multiplier set {0.0018, 0.0035, 0.0067, 0.0130, 0.025, 0.0483, 0.0932, 0.18, 0.36, 0.72, 1.44}.

RD Results

We used the 8 pretrained discrete models for Balle et al., Minnen et al. from compressai as the benchmark.

We re-trained 8 Cheng2020 models on our training dataset from low bitrate and high bitrate (all the models use channel numbers of 192 for comparison).

RD Curve On Kodak dataset with 24 images

Comparison of variable rate methods. The baseline is Balle and use 8 fix-rate models as anchor

Comparison of variable rate methods.The baseline is Minnen and use 8 fix-rate models as anchor

Comparison of variable rate methods.The baseline is Cheng2020 and use 8 fix-rate models as anchor

@INPROCEEDINGS{10222717, author={Tong, Kedeng and Wu, Yaojun and Li, Yue and Zhang, Kai and Zhang, Li and Jin, Xin}, booktitle={2023 IEEE International Conference on Image Processing (ICIP)}, title={QVRF: A Quantization-Error-Aware Variable Rate Framework for Learned Image Compression}, year={2023}, volume={}, number={}, pages={1310-1314}, doi={10.1109/ICIP49359.2023.10222717}}