• Imgcomp
• • Our Contributions
  • File Structure
  • Environment
  • Command
  • Train
  • Test
  • Comparison
  • • Reconstructed image
    • Quality measurements on Kodak
    • • MindSpore version, models trained on imagenet
  • Citation
  • Contributors

Imgcomp

Conditional Probability Models for Deep Image Compression

key words: Image Compression, Conditional Probability Model, Context Compression

The model utilizes the context probability with 3DCNN to get obtain the context to code the latent representation. Also, it utilizes the “mask” mind, namely importance map, for spatial bit-allocation. The paper is published in the CVPR 2018.

Our Contributions

1. Translate from Tensorflow to mindspore
2. Successfully run forward, backward and parameter update
3. Test the mindspore version, and compare the version with the tensorflow one.

File Structure

The translated model file is in imgcomp/demo/model_baseline.py.

demo
├── ckpt
│   └── latest.ckpt      # test model file
├── model_baseline.py    # reproduce model
├── arithmetic_coding.py # arithmetic coding
├── test.py              # test 
└── train.py             # train

Environment

• mindspore-dev==2.0.0dev20230116
  • It is recommended to install through the link.
• cuda 11.1
• python 3.7

Command

Please intall compressai via

pip install -e .

under dictionary CompressAI_MindSpore.

Train

Train the model via:

python train.py -d your/own/dataset/address

Test

We code and decode via:

python test.py -d your/own/dataset/address --real 

We give our trained model in the imgcomp/demo/ckpt, which is trained on the dataset of imagenet. The model can be tested via the before-amentioned. Besides, the test process is so slow that a Kodak image will spend about 1.5h. So we provide the virtual coding version by:

python test.py -d your/own/dataset/address

The instruction will pass the actual encoding-decoding process, which is very close to the real bitrate (<0.1 % difference for any image).

Comparison

Reconstructed image

Mindspore version

bpp:0.630

Official version

Quality measurements on Kodak

MindSpore version, models trained on imagenet

imgname	bpp	PSNR	MSSSIM	run_time
14.png	0.721	21.072	0.958	1.929
11.png	0.63	21.997	0.955	1.232
1.png	0.704	19.157	0.956	1.193
18.png	0.712	20.643	0.955	0.506
6.png	0.655	20.307	0.954	0.313
17.png	0.606	24.307	0.97	1.095
24.png	0.685	19.468	0.963	0.591
20.png	0.538	23.462	0.968	0.66
9.png	0.526	24.839	0.963	1.066
2.png	0.637	24.026	0.939	1.216
13.png	0.791	16.89	0.948	1.173
23.png	0.517	25.718	0.968	0.663
16.png	0.595	23.526	0.959	1.125
7.png	0.548	24.676	0.972	1.07
22.png	0.668	22.753	0.948	0.978
4.png	0.62	24.49	0.955	0.48
8.png	0.707	16.838	0.961	0.876
12.png	0.59	24.8	0.956	0.387
21.png	0.599	21.192	0.959	0.5
10.png	0.564	24.715	0.963	0.714
19.png	0.619	21.539	0.958	0.704
5.png	0.754	19.234	0.965	1.496
15.png	0.6	23.458	0.96	1.149
3.png	0.534	25.244	0.967	1.904
average	0.63	22.265	0.959	0.959

	MSSSIM	bpp
Official TF version	0.987	0.677

Official version can be referred from the link.

Citation

@inproceedings{mentzer2018conditional1,
    Author = {Mentzer, Fabian and Agustsson, Eirikur and Tschannen, Michael and Timofte, Radu and Van Gool, Luc},
    Booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    Title = {Conditional Probability Models for Deep Image Compression},
    Year = {2018}}

Contributors

Name:

Zhuozhen Yu:
email: yuzhuozhen@stu.pku.edu.cn
Please keep free to contact us.

Hua Ye