• e2e_gdn
• • our contibutions
  • file structure
  • environment
  • command
  • test
  • • • use test_ms.py to test mindspore model
      • use test_torch.py to test pytorch model
  • comparision
  • • Quality measurements on Kodak
    • • MindSpore version, models trained on div2k
  • source
  • citation
  • contributors

e2e_gdn

《End-to-end optimized image compression》

key words:image compression, generalized divisive normalization

e2e_gd proposes an image compression method, consisting of a nonlinear analysis transformation, a uniform quantizer, and a nonlinear synthesis transformation. It exhibits better rate–distortion performance than the standard JPEG and JPEG 2000 compression methods.

our contibutions

1.translate from pytorch to mindspore

2.successfully run forward, backward and parameter update

3.test and compare between the mindspore version and the pytorch version with pretrained parameter

file structure

the translated model file is e2e_gdn/model.py

the pretrain paremeter is e2e_gdn_torch/e2e_pretrain.pth for pytorch and e2e_gdn/e2e_pretrain.ckpt for mindspore

environment

• mindspore-dev==2.0.0dev20230116
• ubuntu 16.04
• cuda 11.1
• python 3.7

command

to use the translated model, check

from model import ImageCompressor

model = ImageCompressor

test

use test_ms.py to test mindspore model

cd e2e_gdn
python test_ms.py

use test_torch.py to test pytorch model

cd e2e_gdn_torch
python test_torch.py

comparision

Quality measurements on Kodak

MindSpore version, models trained on div2k

bpp	PSNR	MSSSIM	run_time	GPU Memory(MiB)	lambda
0.447	28.422	0.945	1.227	4004	lambda400
0.697	29.402	0.956	1.04	4004	lambda800
0.967	30.303	0.967	2.514	4004	lambda1500
1.241	30.635	0.973	1.074	4004	lambda3000

source

translated from pytorch

link:liujiaheng/iclr_17_compression: End-to-end optimized image compression (github.com)

citation

@article{balle2016end,
  title={End-to-end optimized image compression},
  author={Ball{\'e}, Johannes and Laparra, Valero and Simoncelli, Eero P},
  journal={arXiv preprint arXiv:1611.01704},
  year={2016}
}

contributors

name: Kaiyu Zheng