PaddleDetection icon indicating copy to clipboard operation
PaddleDetection copied to clipboard

How to train high resolution images?

Open fanyuanxiang opened this issue 2 years ago • 1 comments

问题确认 Search before asking

  • [X] 我已经搜索过问题,但是没有找到解答。I have searched the question and found no related answer.

请提出你的问题 Please ask your question

Training high resolution images, such as 19201080, to detect target object size is basically the same size (300300) , how to adjust the parameters training.

fanyuanxiang avatar Aug 09 '22 07:08 fanyuanxiang

Is the problem about GPU memory or model accuracy?

  • If it is about GPU memory, just set lower batch_size and adjust the learning rate.
  • If it is about model accuracy, in my opinion, the 300*300 size seems well. But you can still refer this config, which is to deal with the small size object and trained with a large input image size.

zoooo0820 avatar Aug 09 '22 09:08 zoooo0820