loss.backward memory

[eric8576]

Hi, my computer GPU is 24GB, but when code run to epoch 2, loss.backward() will show CUDA out of memory.the batch size is one. I do bot know how to fix this! thanks:)

[aaronkujawa]

Hi,
the current training patch size was maximised for a 32GB GPU. You can try to reduce the training patch size, called pad_crop_shape in the code

VS_Seg/params/VSparams.py

Line 76 in 05065cb

self.pad_crop_shape = [384, 384, 64]

and accordingly the sliding_window_inferer_roi_size

VS_Seg/params/VSparams.py

Line 96 in 05065cb

self.sliding_window_inferer_roi_size = [384, 384, 64]

Values should be multiples of 2, for example you can try [256, 256, 64].

[eric8576]

It work!Thank you for your reply!