Correct resizing of images #4791
Replies: 3 comments 1 reply
-
You've chosen to report an unexpected problem or bug. Unless you already know the root cause of it, please include details about it by filling the issue template. |
Beta Was this translation helpful? Give feedback.
-
RectLabel is an offline image annotation tool for object detection and segmentation. |
Beta Was this translation helpful? Give feedback.
-
My guess is that you also need to increase the batch size, otherwise there will not be much gain in reducing the image resolution. If you train on half the image resolution, you can approximately quadruple the batch size (cause GPU memory scales roughly quadratically with the resolution), that should give you a good increase in speed. Keep in mind that you also have to change the learning rate if you change the batch size |
Beta Was this translation helpful? Give feedback.
-
I train on large images (approx. 4000 x 4000 px) and downsize them in a custom mapper. However, this seems to have no effect on the training time. The training time is the same if I do not use the custom mapper.
Instructions To Reproduce the Issue:
I use the following custom mapper:
This is my Trainer train loader:
Expected behavior:
I expected the training time to be reduced. If I replace the images with downscaled images, the training time is significantly reduced: I use a resize factor of .25 and the time is approx. reduced by the same factor. - as expected.
Could someone please tell me how to correctly downsize the images?
Environment:
Beta Was this translation helpful? Give feedback.
All reactions