nnUNet icon indicating copy to clipboard operation
nnUNet copied to clipboard

Published 20 hours ago • verified publisher icon MIC-DKFZ

the problem of loss = nan

Open luoshiyong opened this issue 3 years ago • 1 comments

hi! I want to use nnunet to solve the lits17 dataset ,but the train loss and val loss is always nan,i don't know how to fixd it. Can you give me some suggestions? The following is my process using nnunet

  1. git clone https://github.com/MIC-DKFZ/nnUNet.git cd nnUNet pip install -e .
  2. Convert lits data format: python Task029_LiverTumorSegmentationChallenge.py
  3. Preprocess data: nnUNet_plan_and_preprocess -t 029
  4. Train 2d network: nnUNet_train 2d nnUNetTrainerV2 Task029_Liver 0 --npz

luoshiyong avatar Feb 13 '22 10:02 luoshiyong

Hi, that is strange. I cannot reproduce this problem. Can you please try Task03 from the medical segmentation decathlon? It's the same dataset as LiTS, just from a different source and with some corrected spacings Best, Fabian

FabianIsensee avatar Feb 17 '22 07:02 FabianIsensee

Have you tried the suggested solution? Could you please give us feedback?

Kobalt93 avatar Aug 02 '23 07:08 Kobalt93