why dose the model predict the same label ALL?

[XieXiao1997]

When I try to train a 3D binary classfication refer to 3d_classification\torch\densenet_training_dict, I found that there are always only a few of the same values after a few epoch. So i cheak the ConfuseMatrix ,the predict of the model become odd like this:
Training Set:

Validation Set:

It likes the training make the model using a simple way to classfy only predict the same value.
My trainset/validset‘s labels are evenly distributed. It's hard to understand.
I have tried to reduce the learning rate, add weight-decay in Adam,change the model to Resnet,but all work do not work.

[Nic-Ma]

Hi @demodoom1 ,

Thanks for your experiments here.
I think this example is only to show the functionalities of MONAI components with very few data.
Maybe you can adapt to some real-world dataset for your own requirement.

Thanks.