Open yifeipet opened 2 years ago
ali-zafari
commented
2 years ago Hi Yifei
Yes. the Hyperprior model tries to find a the probability distribution of the latent code and then calculate the entropy of the distribution and report the entropy as the expected bitrate. I am not sure about you second question.
Ali
yifeipet
commented
2 years ago Hi Ali, The predicted entropy is not real bits. Hyper-prior models might have low accuracy when predicting the real standard deviation if the bit rates are low. When the bit rates are very low, a lot of information of original images has been lost so it might not be accurate to use entropy predicted by a hyper-prior model for the real bit rates. So it's not the correct way to use entropy predicted by the hyper-prior entropy estimator for the real bit rates.
If using entropy loss for real bpp, we need to measure the accuracy.
Many people just use entropy loss for real bpp in their papers. But it's not correct since they do not compare the entropy loss predicted by the hyper-prior model with the real bpp. Whether their methods are better than BPG is questionable because the BPG uses real file size but their methods are not in their papers.
I will address this issue in my paper. Thank you!
Sincerely, Yifei
ali-zafari
commented
2 years ago I deeply appreciate your quick response yifeipet. I am going to discuss my understandings from reading papers in this area, and I would be really grateful if you could share your thought anywhere I am wrong.
The predicted entropy is not real bits
Sure. No one uses the entropy predicted by the hyper-prior as the real bit rate. It is just an approximation to enable you optimizing your model on the rate distortion trade-off.
Many people just use entropy loss for real bpp in their papers.
As far as I have read papers in neural image compression area, when it comes to compare performance with BPG, JPEG and any other codec, everybody report the actual bitrate which is calculated from the bitstream size of the output of the entropy coder (like arithmetic coder). I totally agree that comparing the entropy loss predicted by the hyperprior model with the final bitrate of BPG is obviously wrong.
Regards, Ali
Justin-Tan
commented
2 years ago Hi all,
Sorry for the slow response. I think the theoretical bpp is computed as the entropy of the latents and hyperlatents as output by their respective probability models.
On Mon, Mar 21, 2022 at 7:17 PM Ali Zafari @.***> wrote:
I deeply appreciate your quick response yifeipet. I am going to discuss my understandings from reading papers in this area, and I would be really grateful if you could share your thought anywhere I am wrong.
The predicted entropy is not real bits
Sure. No one uses the entropy predicted by the hyper-prior as the real bit rate. It is just an approximation to enable you optimizing your model on the rate distortion trade-off.
Many people just use entropy loss for real bpp in their papers.
As far as I have read papers in neural image compression area, when it comes to compare performance with BPG, JPEG and any other codec, everybody report the actual bitrate which is calculated from the bitstream size of the output of the entropy coder (like arithmetic coder). I totally agree that comparing the entropy loss predicted by the hyperprior model with the final bitrate of BPG is obviously wrong.
Regards, Ali
— Reply to this email directly, view it on GitHub https://github.com/Justin-Tan/high-fidelity-generative-compression/issues/41#issuecomment-1074320046, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGRNY6B5UZ4JKNT4HXMD7XLVBDDLVANCNFSM5OE6CZSQ . You are receiving this because you are subscribed to this thread.Message ID: <Justin-Tan/high-fidelity-generative-compression/issues/41/1074320046@ github.com>
yifeipet
commented
2 years ago Hi all, Sorry for the slow response. I think the theoretical bpp is computed as the entropy of the latents and hyperlatents as output by their respective probability models. … On Mon, Mar 21, 2022 at 7:17 PM Ali Zafari @.***> wrote: I deeply appreciate your quick response yifeipet. I am going to discuss my understandings from reading papers in this area, and I would be really grateful if you could share your thought anywhere I am wrong. The predicted entropy is not real bits Sure. No one uses the entropy predicted by the hyper-prior as the real bit rate. It is just an approximation to enable you optimizing your model on the rate distortion trade-off. Many people just use entropy loss for real bpp in their papers. As far as I have read papers in neural image compression area, when it comes to compare performance with BPG, JPEG and any other codec, everybody report the actual bitrate which is calculated from the bitstream size of the output of the entropy coder (like arithmetic coder). I totally agree that comparing the entropy loss predicted by the hyperprior model with the final bitrate of BPG is obviously wrong. Regards, Ali — Reply to this email directly, view it on GitHub <#41 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGRNY6B5UZ4JKNT4HXMD7XLVBDDLVANCNFSM5OE6CZSQ . You are receiving this because you are subscribed to this thread.Message ID: <Justin-Tan/high-fidelity-generative-compression/issues/41/1074320046@ github.com>
Hi Justin, Which bpp do you think they report on their paper? Theoretical bpp or real bpp? When I trained the GAN using your codes, I found there was a big gap between real bpp and theoretical bpp. I set the target bpp to 0.07 bpp. But the real bpp is like 2 times of theoretical bpp.
Thank you!
Sincerely, Yifei
Hello Justin, Thank you for your outstanding reference codes and I learned a lot. I got high-quality reconstructed images. I have a question about "theoretical bpp". Is it calculated by the hyperprior model? Is the hyperprior model accurate because I see the big gap between real bpp and "theoretical bpp"?
Thank you!
Sincerely, Yifei