search

ljwztc / CLIP-Driven-Universal-Model

[ICCV 2023] CLIP-Driven Universal Model; Rank first in MSD Competition.
Other
521 stars 58 forks source link

question about implementation details #34

Closed ShawnHuang497 closed 9 months ago

ShawnHuang497 commented 10 months ago

image Hi, In your paper u said that u use batch size 6 with a patch size of 96 × 96 × 96 per device NVIDIA RTX A5000, which is with 24G video memory. But I use batch size 2 with a patch size of 96 × 96 × 96 per device NVIDIA RTX A5000, encountering cuda is out of memory. And I tried three backbones that were unet, swinunetr, unetpp, encountering the same issue.

ljwztc commented 10 months ago

There is a parameter --num_samples that also effect the GPU consumption. So, the patch number in each card is num_samples * batch_size. In our experience, we can train the model in 24G card with 2 * 1 or 1 * 2.