inofficialamanjha
I have come across a problem, when the ration is set as 0.5
Hello， I have run your code CSNET， Thank you for your work. But I have some questions: when the ratio =0.01, 0.05, 0.1, the code can achieve the same performance as papers, at the same time when the ration is higher such as 0.5, the psnr results are closer to 35.5+, a little lower than 40+db in the paper。 Maybe somewhere I made mistakes you know, or you have the same problems we can discuss to solve it. Looking forward to your reply.
Yeah sure!
Let's connect on google meets. Convey the appropriate time as per Location. ( Also send a google meet link for the same )
Kindly, prefer keeping it between 13:00 to 20:00 ( IST ).
It's been a bit of time since I have spent some time, working on it - but happy to discuss the problem and provide inputs.
Thanks & Regards, Aman Jha.
hhh Thanks for your reply. You are kind and helpful. But my English is a little poor maybe can not get smooth communication with you and may waste a lot of your precious time. May be we can discuss by email or just here. Thank you very much. My email address: [email protected] The question is that can your test result when the ratio is set as 0.5 can achieve the psnr results in the paper?
My results shows that when the ratio gets higher, the differences between my results and the paper is larger.
I'm confused about this results for one week, so I tried to ask for help. Thank you very much again.
Okay, I will definitely look into the issue and respond back to you by 2nd August. ( will try to look into it ASAP )
i will try my best to find the solution If i have some ideas, I'll leave notes here
hey Zhouziqun, thank you for waiting.
However, I was unable to find out a viable solution for the problem.
Were you able to solve it ?
I have made some changes
firstly, I changed the channel of the net ( one channel of Y -> three channels of RGB and then get the channel Y through the three channels, finally I goy PSNR )
secondly, I re-read the paper i find this net has two losses (the one loss is calcuated as the diffenence between the init-reconstruction image and the label, the other one is calcuated as the difference between the final image and the label)
Last but not least
i trained and tested subrate 0.01 0.3 0.5 to compare. I show you my results and the change, maybe you can continue to improve on my work.
I'm not sure if my English ability is enough to express clearly. Thank you very much for your willingness to help me solve the problem. I look forward to you working with me to continue to solve this problem. Thank you very much
Don't worry about your communication - it would be the least we are concerned with.
Thanks for the feedback, will try to respond back to you quickly this time - let me re-read & reconstruct the whole project