search

avBuffer / UNet3plus_pth

UNet3+/ UNet++/UNet, used in Deep Automatic Portrait Matting in Pytorth
233 stars 39 forks source link

hi train with my own data #1

Open zhangyunming opened 4 years ago

zhangyunming commented 4 years ago

i train the v3 with my own data that with lables 0:backgrounbd and 1:portrait, just two classes ,so the n_classes=1 , but you masks labels are 1-255 , and my train loss is not change and the test result are all black 0 . so how can i do .thanks

zhangyunming commented 4 years ago

the bug is ok.its similar with pytorch-unet

avBuffer commented 4 years ago

hey, you maybe change into UNet2plus to try. If the error is same, there is wrong in your data or using loss function. Else there is wrong in UNet3plus model structure.

avBuffer commented 4 years ago

Hey,thanks! I tried it and found the training loss was changed after about 18 steps. In fact, from step-0 to step-17, the training loss had a bit a bit changed. You maybe train a long steps.

jialeqaq commented 1 year ago

I train myself on data, DICE is low and constant, how do I solve it

jialeqaq commented 1 year ago

i train the v3 with my own data that with lables 0:backgrounbd and 1:portrait, just two classes ,so the n_classes=1 , but you masks labels are 1-255 , and my train loss is not change and the test result are all black 0 . so how can i do .thanks

hello，why my training data DICE is very low and the more I train, the worse it gets