search

ge-xing / Diff-UNet

Diff-UNet: A Diffusion Embedded Network for Volumetric Segmentation. (using diffusion for 3D medical image segmentation)
Apache License 2.0
155 stars 21 forks source link

Training results using 2 GPUs #18

Open jhonson592 opened 1 year ago

jhonson592 commented 1 year ago

Thanks for releasing the paper and the source code!

I experimented with GPU=2 since I only have 2 GPUs. I experimented with batch_size=4 and found that the mean_dice during training was abnormally low compared to other people's results.

Results of the experiment with batch_size=4 training mean_dice = 0.5956 wt = 0.9142 tc = 0.6453 et = 0.6193

mean_dice train_loss et

Are there any parameters or codes that should be changed by reducing the number of GPUs?

920232796 commented 1 year ago

The result is strange. In my experiment, the result is not related to the number of gpus.

jhonson592 commented 1 year ago

Thank you for reply!

Maybe there is a difference between the different versions. If you don't mind, could you please tell me the versions of python, torch, numpy, monai and SimpleITK?

Liour commented 12 months ago

Thank you for reply!

Maybe there is a difference between the different versions. If you don't mind, could you please tell me the versions of python, torch, numpy, monai and SimpleITK?

may I ask about your Python, Torch, and monai versions, I encountering many warnings when training start, and I cannot achieve the performance you reproduced.