search

YaoleiQi / DSCNet

Pytorch Implement of Dynamic Snake Convolution (ICCV2023)
418 stars 37 forks source link

OOM！ #32

Open Itsanewday opened 11 months ago

Itsanewday commented 11 months ago

Thanks for sharing! When I apply DSC3D block in my own net and dataset. I meet the OOM problem, the training patch_size is 128x128x128, the batchsize is 2 and the input_channels is 32. How can i sovle this problem？ Or there are any tricks to reduce the GPU Memroy? BTW, my GPU is GTX 3090 with 24GB memory.

Li-Yicun commented 9 months ago

缩小patch_size

YaoleiQi commented 6 months ago

Thank you very much for your attention. Due to recent changes in my work, I couldn't reply promptly. During training in 3D, we set the batch size to 1. Of course, there's a trick: you can incorporate downsampling in the input part and upsampling before the output. Additionally, it's possible to optimize memory consumption by borrowing from parts of DSC2D_pro.