Training analysis: not able to overfit a training set. Details provided, need helps

Hi everyone,

I’m trying to analyze a training session of my own, and I need some helps with the issue of not able to overfit the training set (which supposed to be overfitted if we keep training more and more in general).

Here are the details.

  • fastai2 version: 0.0.13
  • Problem: Segmentation (building footprints).
  • Dataset: SpaceNet Buildings v2.
  • Training template: I follow the standard Camvid tutorial. You can see more details in the executing notebook here
  • Issue: the training loss always blows up after 1 or 2 epoch (over 5 in total).
  • Things I have tried but did not solve the issue: very low learning rates, weight decay=0.
  • Plot:

Anyone’s thoughts are welcome.