Open wzmw-zr opened 1 year ago
I have also discovered this problem. Have you solved it?
I have also discovered this problem. Have you solved it? After 24 hours of training (DETR) on 8*2080Ti, the memory useage was over 400GB!
I have also discovered this problem with pytorch version 2.0. Have you solved it? After 24 hours of training (DETR) on 8*2080Ti, the memory useage was over 400GB!
When I use pytorch version 1.13, the memory doesn't overflow anymore.
WOW!! thank you @mypydl !!
You're a life saver!! :D
我发现了是RandomCrop的问题,类detr中数据增强使用了RandomCrop会引起CPU内存泄露,我去掉就没有这个情况了。
When I train DETR-like object detectors (e.g. DETR, DINO...) in mmdetection3.0, the occupied memory of RAM will increase fast, so the training process will be killed when there is no free space in RAM. However, when I switch to mmdetection2.52.2, the occupied memory of RAM will increase slowly.
In mmdetection2.52.2, the RAM usage and other information in the training process of DETR are as follows:
In mmdetection3.0.0, the RAM usage and other information in the training process of DETR are as follows: