The training and validation loss dropped to -0.6, and it stopped dropping

[MOMOANNIE]

Hello FabianIsensee,
I use nnunet to train my own data. After 1000 epochs of training, my training loss and validation loss are both -0.6+, and they don't continue to decrease. What's going on?

Below is my training process：

[dojoh]

Hello,

sorry for the late response!
Is this question still relevant? To me it looks the loss might still decrease if training is continued. In what way did the loss no further decrease? Did you at any point run e.g. 2000 epochs?

Cheers
Ole

[MOMOANNIE]

Is this question still relevant

Yes, this also happens when I train with nnunetV2

Did you at any point run e.g. 2000 epochs

I have tried to modify the number of epochs, but failed, it is still 1000, what needs to be modified to accurately modify the number of epochs?

[dojoh]

There are a couple of predefined trainers with more epochs, see https://github.com/MIC-DKFZ/nnUNet/blob/b4e97fe38a9eb6728077678d4850c41570a1cb02/nnunetv2/training/nnUNetTrainer/variants/training_length/nnUNetTrainer_Xepochs.py

You can invoke these trainers using the -tr flag, e.g. nnUNetv2_train DATASET_NAME_OR_ID UNET_CONFIGURATION FOLD -tr nnUNetTrainer_8000epochs

[975827738]

Is this question still relevant

Yes, this also happens when I train with nnunetV2

Did you at any point run e.g. 2000 epochs

I have tried to modify the number of epochs, but failed, it is still 1000, what needs to be modified to accurately modify the number of epochs?

Is this question still relevant

Yes, this also happens when I train with nnunetV2

Did you at any point run e.g. 2000 epochs

I have tried to modify the number of epochs, but failed, it is still 1000, what needs to be modified to accurately modify the number of epochs?

Hello,
Is it normal for loss to be a negative number

[iWangTing]

这个问题仍然相关吗

是的，当我使用 nnunetV2 训练时也会发生这种情况

您是否在任何时候运行过，例如 2000 个时代

我试过修改纪元数，但失败了，还是1000个，需要修改什么才能准确修改纪元数？

这个问题仍然相关吗

是的，当我使用 nnunetV2 训练时也会发生这种情况

您是否在任何时候运行过，例如 2000 个时代

我试过修改纪元数，但失败了，还是1000个，需要修改什么才能准确修改纪元数？

你好 损失为负数正常吗

这个问题仍然相关吗

是的，当我使用 nnunetV2 训练时也会发生这种情况

您是否在任何时候运行过，例如 2000 个时代

我试过修改纪元数，但失败了，还是1000个，需要修改什么才能准确修改纪元数？

这个问题仍然相关吗

是的，当我使用 nnunetV2 训练时也会发生这种情况

您是否在任何时候运行过，例如 2000 个时代

我试过修改纪元数，但失败了，还是1000个，需要修改什么才能准确修改纪元数？

你好 损失为负数正常吗

这个问题仍然相关吗

是的，当我使用 nnunetV2 训练时也会发生这种情况

您是否在任何时候运行过，例如 2000 个时代

我试过修改纪元数，但失败了，还是1000个，需要修改什么才能准确修改纪元数？

这个问题仍然相关吗

是的，当我使用 nnunetV2 训练时也会发生这种情况

您是否在任何时候运行过，例如 2000 个时代

我试过修改纪元数，但失败了，还是1000个，需要修改什么才能准确修改纪元数？

你好 损失为负数正常吗

I have also encountered the situation that loss is negative, may I ask if you have solved it? What is the basis for a negative loss value?

[ravikumawat7716]

@FabianIsensee , @dojoh I have the same question, why it is showing the loss in negative.

[iWangTing]

@FabianIsensee , @dojoh I have the same question, why it is showing the loss in negative.，我有同样的问题，为什么它以负数显示损失。
Hi，
I have realized the problem.
The loss function of nnUNet is designed in such a way that the closer the value is to -1, the smaller the model's loss.

[e-FunJs]

@FabianIsensee，@dojoh 我也有同样的疑问，为什么显示负数的损失。
For nnUNet, the closer the loss value is to -1, the better the performance is.