训练loss从十几或者二十几轮开始变为nan，并且dice也在逐渐下降

[zhanshenhhh]

您好，我在训练时按照您所给的配置进行训练，但是在训练十几或者二十几个epochs后，损失值就突然变成nan并且dice指数也在逐渐下降，请问您知道是什么原因吗？

[tiffany9056]

您好，我没有遇到这个情况，但是您可以尝试调整您的learning rate看是否会有差

[Mr-wan123]

我在复现的过程中也遇到了这个问题，调整了学习率和SAM的参数还是不行，不知道现在有没有什么解决方法呢Duplicate of #

[renlyly]

Hi, thank you for your work!

I'm currently trying to reproduce your results, but I ran into the same issue mentioned above. I’ve tried adjusting the learning rate and the SAM parameters, but the problem persists.

Also, I encountered compatibility issues when installing uumamba with default PyTorch 2.0.1, so I had to upgrade to 2.1.2. Which version of Mamba-SSM is compatible or recommended?

Additionally, I noticed that uumamba refers to nnU-Net version 2.3.2, but I couldn’t find this version on the official nnU-Net GitHub repository. In order to run the processing and training step, an older version of nnU-Net is needed to run it successfully. I am not sure if this is the reason

Thanks again for your contribution!