search

MrGiovanni / UNetPlusPlus

[IEEE TMI] Official Implementation for UNet++
Other
2.3k stars 541 forks source link

Getting very bad validation metrics #67

Open kirillmeisser opened 3 years ago

kirillmeisser commented 3 years ago

Hi, I have trained the U-Net++ on my 2D image dataset on all the five folds and I somehow get very bad metrics on the validation sets. For example, the mean Jaccard index for the validation set of fold 0 is 0.0761 and the other folds are also more or less there. The network does not seem to learn and I do not understand why. A couple of weeks ago I trained the original nnUNet from FabianIsensee and everything worked fine and I got really great results. I did everything basically in the same way here but without success. Did anyone encounter this issue? Does one have to change something when using 2D data? Any help is greatly appreciated!

mueller-franzes commented 2 years ago

Hi @kirillmeisser, I am facing the same problem. Have you found a solution?

aadya940 commented 2 years ago

Yes, I'm encountering the same error