Training a network for segmentation doing very badly

Part 1 (2019)
1

I wanted to try fastai on a simple segmentation problem, and tried it on a previous Kaggle competition. However, the Dice value is quite poor while the loss is low. What am I doing wrong?

Here’s a link to my kernel

2

Possibly nothing. Run it for longer. Long enough that the validation error isn’t falling AND the dice isn’t rising. Segmentation kernels can (sometimes) take hours to run. The loss may look low, but it may go much lower, that’s why the metric is important.

3

Thanks… Looking at the kernel again, I realized that the fit_one_cycle function is running for 10 epochs, it just doesn’t display as such. I plotted the losses and metrics and it seems the loss won’t decrease any further (see below), and the Dice score is up to 0.6, which is OK but I would definitely want it to improve. I haven’t unfreezed the model yet, but should I train the frozen model even more or is this a good place to stop?

image

The kernel is the same

4

@digitalspecialists I unfroze the network and did not see significant improvement… Any recommendations?

(note, increasing batch size or image size cannot be done because of memory limitations of the Kaggle Kernel)