The issue of CUDA memory usage during training

[Li-4]

Firstly, thank you for your outstanding work and contribution. I found that when training on the 3D dataset of 801Liver, the memory usage was very high, reaching 29140 MiB. May I ask if this is normal？

[eclipse0922]

I'm not the author, but for reference The current implementation does not match the paper.

6

[MrBlankness]

Firstly, thank you for your outstanding work and contribution. I found that when training on the 3D dataset of 801Liver, the memory usage was very high, reaching 29140 MiB. May I ask if this is normal？

Perhaps this is an inherent issue with nnU-Net. If it's truly impossible to meet the memory requirements, we suggest using only the network model itself that we proposed, rather than employing the associated training pipeline.

[xwgit2023]

请问有在二维数据集上跑，图片尺寸是多少，我发现256*256特别慢