Open Feynman1999 opened 10 months ago
Hi, If I remember correctly, I had done such experiments by removing predictive module and learning z itself. I even tried with auxiliary loss function but the performance was just not as good as the presented one. The intuition is the latent variable z might actually carry some useful information needed to reconstruct the images/video. Due to the special structure of Invertible neural network, this z can actually be whatever the model likes it to be. This is my understanding and it has been a long time since I finished the research. There might be another new researches going on. Hope this help.
thank you for reply! It seems that there is very little work related to video rescaling? Do you think there is still room for research in this direction?
By the way, is there no checkpoint available? It would be very helpful for me to have a pre trained model
Frankly speaking, the image/video rescaling task might not worth digging in. But it has great potential to extend to various tasks such as image/video compression. Honestly, I haven't been doing this line of research for a while so this is just my humble thought.
About checkpoints, sorry that I have no access to my school account anymore. I don't have copy either.
FYI,
If you really need the checkpoints, my lab might have a copy. Please contact http://mapl.nctu.edu.tw/
On Tue, Sep 12, 2023 at 10:28 Yuxiang Chen @.***> wrote:
By the way, is there no checkpoint available? It would be very helpful for me to have a pre trained model
— Reply to this email directly, view it on GitHub https://github.com/ding3820/MIMO-VRN/issues/7#issuecomment-1714871222, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGZQSJPSUL3XF37A3LWLSETXZ7CGDANCNFSM6AAAAAA4TFP3BE . You are receiving this because you commented.Message ID: @.***>
图像rescaling任务中,一般没有这个predictive module,请问这里如果不使用这一结构,还是sample from normal distribution,让网络自己学z的分布,结果如何?期待您的指点~