Train my own dataset about Resize problems.

Open freeforhub opened this issue 10 months ago • 4 comments

Hello, thank you and your team for your contribution! When I want to train my own data set (the size is 3072X2048), I adjusted the configuration file you mentioned (REsize to 1200x1600, batch size: 1), but I will get an RuntimeError: The size of tensor a (625) must match the size of tensor b (3750) at non-singleton dimension 1. I can train normally after changing in RE-DETR. Can you tell me what configuration files I need to modify to train my own data set? Thank you, good wishes!

Apr 01 '25 03:04 freeforhub

Excuse me, does BatchImageCollateFunction only support input with consistent height and width? So how should I set my image?

Apr 01 '25 13:04 freeforhub

If I'm not wrong, the image height and width should be divisible by 32.

Apr 03 '25 02:04 dnth

I tried, but I couldn't. I can train it by padding it to square data, but will 0 pixel filling lead to model learning deviation?thanks！

May 08 '25 05:05 freeforhub

I'm having the same problem. Please let me know if there's a solution.

Jun 10 '25 15:06 snoopybingo