search

Amshaker / unetr_plus_plus

[IEEE TMI-2024] UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation
Apache License 2.0
340 stars 32 forks source link

train_loss #71

Closed WangXiaoXv closed 4 months ago

WangXiaoXv commented 5 months ago

@Amshaker Hello, I apologize for any inconvenience caused. When running the training_scripts/run_training_acdc.sh script, I encountered a negative train_loss. I would like to know the reason behind this issue. I am eagerly awaiting your response! 图片1 ![Uploading 屏幕截图 2024-03-22 202334.png…]()

Amshaker commented 4 months ago

The loss is based on Dice + CE.

Please check this issue, it is the as in nnFormer, which our codebase is based on: https://github.com/282857341/nnFormer/issues/40

We should add a constant value to the dice, but here it's ignored.