UNET (Lesson 14) With other datasets

Hello everyone,

I have a newbie question, I think. I watched some of the first part videos and the last video about image segmentation and I want to thank the creators of fastai because after trying to understand everything from other normal courses I was feeling pretty lost about deep learning.

Now to the point. I have a segmentation problem aerial images that I need to solve and really very little VOC ground truths but some raw data (right now ~500 marked images but with the possibility to make more). I wanted to practice detecting one class, trees in this case. But this error appears whenever I try to reuse the carvana example (that works perfectly for me).

I have read other answers on the forum that deal with this problem, but it is on the fit phase that it crashes, not on the best learning rate finder, (if I remove the best learning rate finder it still crashes on the fit phase) with 2 possible explanations:

  1. not enough data (possibly true for me too)
  2. and to try to change the cycle_len to something > 1

But none of this changes work. Any recommendations? Or I start to put more effort on labeling and then creating ground truths.

if I remove the last 2 variables :cycle_len = 100, use_clr = (5, 8). I end up with:

Thanks for listening and the help.

I solved the problem about the same week. For anyone needing it in the future, it was due to :

val_idxs = list(range(CURRENT_TOTAL_DATA)

Being given more data than there was in total, which is of course wrong in 2 levels:

  1. This is for a validation dataset (meaning is not right at all!). Use only a % of your total data.
  2. It accepts CURRENT_TOTAL_DATA - 1.