About:the loss function and the value of lambda

zhang lu

unread,

Jul 28, 2021, 4:11:53 AM7/28/21

to tensorflow-compression

Hi Johannes,

And I have another problem about the artical VARIATIONAL IMAGE COMPRESSION WITH A SCALE HYPERPRIOR. If I will optimized for MS-SSIM,what the loss function I will use? I know the loss=λ*(1-D)+R,but how could I get the 'D'?

When I used the train_msssim=tf.reduce_mean(tf.image.ssim_multiscale(x_tilde,x,255)) to get the average,the result is not well. Is the loss function wrong or the value of λ is wrong? Are the values of λ to optimized for MS-SSIM same as optimized for MSE?

Very Very looking forward your reply!Thanks for your time.

Best wishes,

ZhangLu

Benoit Brummer

unread,

Jul 28, 2021, 4:16:47 AM7/28/21

to tensorflow-compression

The lambda values are indeed different and much smaller (at least in the pytorch implementation). I can optimize for the MS-SSIM by pre-training with MSE and fine-tuning a little bit. The rate-distortion degrades if I try to optimize for too long and the results are off if I train directly with MS-SSIM only. I'm also interested in how to make this work better.

zhang lu

unread,

Jul 28, 2021, 10:52:44 PM7/28/21

to tensorflow-compression

Sincerely thank you for your reply.

How to get he artical's result of optimised for MS-SSIM? And how to optimize for the MS-SSIM by pre-training with MSE and fine-tuning a little bit?
Looking forward to your reply.

zhang lu

unread,

Jul 28, 2021, 10:58:14 PM7/28/21

to tensorflow-compression

And the compressai had loss function of optimized for MS-SSIM,which is loss=λ*(1-D)+R, Is 'D' not only the ms-ssim?

On Wednesday, July 28, 2021 at 4:16:47 PM UTC+8 trou...@gmail.com wrote:

Reply all

Reply to author

Forward