Closed qsunyuan closed 1 year ago
No worries about the memory increase as the RPN are trained to recall regions of interest with higher confidence scores. However, half the learning rate and double the training iteration number as you use 8 GPUs instead of 16.
Excellent work!
I run ov-coco benchmark using the following command on 3090 x 8 with cuda 11.3 and torch1.12.1:
but the GPU memory keep increase during training. Any suggestion?