flyywh / Neural-Syntax-Code

Neural-Syntax

Code for CVPR 2022 Paper "NEURAL DATA-DEPENDENT TRANSFORM FOR LEARNED IMAGE COMPRESSION"

Author: Dezhao Wang, Wenhan Yang, Yueyu Hu, Jiaying Liu

Arxiv Link: https://arxiv.org/abs/2203.04963

Project Page: https://dezhao-wang.github.io/Neural-Syntax-Website/

Evaluation

• Here we provide our pretrained model optimized for MSE at lambda 0.0015 for reference.

python eval.py --data_path {test_set_path} --lambda {lambda_value} --weight_path {tested_checkpoint_path} [--tune_iter {pre_prpcessing_tune_iteration_num}] [--post_processing] [--pre_processing] [--high]
e.g. python eval.py --data_path ../Kodak/ --lambda 0.0015 --weight_path ./weights/mse0.0015.ckpt --post_processing --pre_processing

Training

• We use a two-stage training strategy.

  • In the first stage, we train the transform model and entropy model.

  • In the second stage, we train the post-processing model with other modules fixed.

python train.py --train_data_path {"train_set_path/*"} --lambda {lambda_value} --checkpoint_dir {saved_checkpoint_dir}  [--weight_path {pretrain_model}] [--batch_size {batch_size}] [--lr {learning rate}] [--post_processing] [--high]
e.g. python train.py --train_data_path "../DIV2K_train_HR/*png" --lambda 0.0015 --checkpoint_dir ./weights/  --weight_path ./weights/mse0.0015.ckpt

Acknowledgement

We implement our post-processing model with the help of https://github.com/wwlCape/HAN.