search

yuhongtian17 / Spatial-Transform-Decoupling

MIT License
62 stars 5 forks source link

Memory Problems #17

Open LightZH opened 1 week ago

LightZH commented 1 week ago

How much video memory is required for each of these four configuration training: STD with Oriented RCNN and ViT-B、STD with Oriented RCNN and HiViT-B、rotated_imted_faster_rcnn_vit_small_1x_dota_le90_8h_stdc_xyawh321v、rotated_imted_oriented_rcnn_vit_small_1x_dota_le90_8h_stdc_xyawh321v

I have been encountering video memory problems with RTX 3090 when training: the set samples_per_gpu=1, workers_per_gpu=2.

yuhongtian17 commented 1 week ago

If "CUDA out of memory" happens, make all "use_checkpoint=True" instead of "False". ViT-S needs 11GB at least and ViT-B/HiViT-B needs 23GB at least with our RTX 4090.