Closed pkqbajng closed 5 months ago
Hello, the PaSCo(M=1) model is relatively lightweight and can be trained on a v100 GPU. However, the ensemble version is significantly heavier. For instance, with 3 subnets, it requires approximately 44GB of memory on average (although I may not recall the exact figure since we made some minor design improvements after submission that improved memory consumption during training). Nonetheless, this number can fluctuate considerably due to variations in the number of voxels when combining different scans.
Hi, sir, thanks for your excellent job. I found that PaSCo adopts a batch size of 2 on 2 A100 gpus. How much the memory of GPU is costed on a single GPU?