在synapse数据集上loss为nan

[Bougaivillea10]

作者你好，我按照你说的讲vmunet项目中的vmunet.py更换为UltraLight_VM_UNet.py，但是训练的loss一直为nan，你有遇到过类似的问题吗？有什么解决的建议吗？感谢您的工作。

[wurenkai]

Hi, I think it would be worth checking if the output 'return torch.sigmoid(out0)' of the model 'UltraLight VM-UNet' is changed to 'return out0' and if Loss is correctly configured to accommodate multi-objective segmentation.

[Bougaivillea10]

Hi, I think it would be worth checking if the output 'return torch.sigmoid(out0)' of the model 'UltraLight VM-UNet' is changed to 'return out0' and if Loss is correctly configured to accommodate multi-objective segmentation.

Thanks for your responding. I have changed 'return torch.sigmoid(out0)' into 'return out0', and chose CeDiceLoss. However, it doesnt fix the problem. Looking forward to your further reply.

[ruti0823]

@Bougaivillea10 Hello, I also encountered such a problem, can you solve it?

[Bougaivillea10]

@Bougaivillea10 Hello, I also encountered such a problem, can you solve it?
Are you still working on this project? I wonder if I can discuss it with you. My email is peiwenhuang1011@qq.com