DVC_P
《DVC-P: Deep Video Compression with Perceptual Optimizations》
This work is based on OpenDVC (an open source Tensorflow implementation of DVC), but improves it with perceptual optimizations (i.e., a discriminator network and a mixed loss are employed to help our network trade off among distortion, perception and rate, and nearest-neighbor interpolation is used to eliminate checkerboard artifacts).
our contibutions
1.translate from tensorflow to mindspore
2.successfully run forward, backward and parameter update
3.test performance in mindspore version
environment
- mindspore-dev==2.0.0dev20230109
- ubuntu 16.04
- cuda 11.1
- python 3.7
command
please intall compressai via:
cd ./CompressAI_MindSpore
pip install -e .
train
Our train experiments are conducted on Vimeo-90k dataset.
python train.py --alpha 256/512/1024/2048
test
Our test experiments are conducted on UVG test dataset.
python run_test.py
mindspore performance
The results of different bitrate models on UVG test dataset:
lambda |
bpp |
PSNR |
MSSSIM |
time_cost |
GPU M |
256 |
0.109 |
33.897 |
0.924 |
4883 |
8370 |
512 |
0.136 |
35.104 |
0.937 |
4989 |
8370 |
1024 |
0.175 |
36.284 |
0.947 |
5082 |
8370 |
2048 |
0.254 |
37.157 |
0.955 |
4968 |
8370 |
source
translated from tensorflow
link: DVC_P-tensorflow
citation
@article{DVC-P,
title={DVC-P: Deep Video Compression with Perceptual Optimizations},
author={Saiping Zhang and Marta Mrak and Luis Herranz and Marc Gorriz Blanch and Shuai Wan and Fuzheng Yang},
journal={arXiv preprint arXiv:2109.10849},
year={2021}
}
contributors
name: Kaiyu Zheng, Yongchi Zhang