-
Notifications
You must be signed in to change notification settings - Fork 10
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Why don't use large patch #7
Comments
Hi @DEAN2012-W, thanks for your interest in your work. Yes, I agree with you large patch size can solve the mosaic edge problem. However, when the patch size is larger, there will be much greater GPU consumption since self-attention demands more GPU memory in the backpropagation. Even armed with the window attention as proposed in Swin transformer (SwinIR in deonising), a powerful GPU like 3090 can barely handle images of 512*512 with a batch size of 1. |
Also, smaller patch size allows larger batch size for training which indicates more diverse samples and more training stability. |
Thanks,your answer is really helpful for me! |
My pleasure!
获取Outlook for Android<https://aka.ms/AAb9ysg>
…________________________________
From: DEAN2012-W ***@***.***>
Sent: Saturday, June 25, 2022 11:55:03 PM
To: wdayang/CTformer ***@***.***>
Cc: dayang_wang ***@***.***>; Comment ***@***.***>
Subject: Re: [wdayang/CTformer] Why don't use large patch (Issue #7)
Thanks,your answer is really helpful for me!
―
Reply to this email directly, view it on GitHub<#7 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AFQCHCXYDX62R6NT5ZXVZW3VQ7IBPANCNFSM5Z3G53EQ>.
You are receiving this because you commented.Message ID: ***@***.***>
|
Hello,I have just read your paper,the idea of Inference of the CTformer to overcome the mosaic edge of the patch,.
I have to say it's really a good ides,but I think what about if i set the patch equal to image size((also it's resule in large Flops) alse can overcome this problem.
The text was updated successfully, but these errors were encountered: