Co-DETR
Co-DETR copied to clipboard
5 RTX4090s are still not enough for this project?
I only used the co_dino_5scale_swin_large_3x_coco.py config which is smaller, and has resized my pictures to 512*512, with 5 RTX4090 GPUs to train. However the memory was still not enough. Is that normal? Do the config files in this project support mixed precision training? I have added this line into the config trying to reduce memory consumed, but it seemed no use: fp16 = dict(loss_scale=512.).
- Please show me your training config.
- This repo does not support fp16 training.