search

LeeJunHyun / Image_Segmentation

Pytorch implementation of U-Net, R2U-Net, Attention U-Net, and Attention R2U-Net.
2.72k stars 600 forks source link

Class-balanced CE loss function leads to worse result? #10

Closed rosefun closed 5 years ago

rosefun commented 5 years ago

I find that the class-balanced cross entropy loss function leads to smaller loss, about 0.001 - 0.008 in my dataset, and the performance is worse than using Iou as loss function which has bigger loss value .

LeeJunHyun commented 5 years ago

Hi, @rosefun . I've never used class-balanced CE loss function, and also I don't know what is your dataset. So I have no idea about that :( Can you check class distribution of your dataset?

LeeJunHyun commented 5 years ago

If you don't have further issues, I will close the issue. I hope my answer was helpful :)