0
$\begingroup$

I am developing fully convolutional model for semantic segmentation task and I tried to use spatial dropout layers to prevent overfitting of my model.

My model has interesting learning curves and I am not sure if it is overfitting or not. According to loss function my model is overfitting because it has smaller train loss than validation loss. But according to graph of mean intersection over union (mIoU) or other metrics my model is not overfitting because train metric curves (dashed) are uniformly under the validation metric curves (solid). How can I now tell if my model is overfitting or not?

Here is graph. In left image is loss function and in right one are metrics (jaccard is mIoU, dice is F1 score): enter image description here

Improve this question
$\endgroup$
10
  • $\begingroup$ It's probably overfitting. $\endgroup$
    – user2974951
    Commented Jun 10, 2021 at 12:26
  • $\begingroup$ You'll pretty much always have a higher out-of-sample loss than in-sample. That should not concern you. $\endgroup$
    – Dave
    Commented Jun 10, 2021 at 12:26
  • $\begingroup$ Ok so when someone is telling about overfitting, it's only according to loss function? $\endgroup$
    – Many
    Commented Jun 10, 2021 at 12:29
  • $\begingroup$ I would expect that loss is somehow correlated to metrics $\endgroup$
    – Many
    Commented Jun 10, 2021 at 12:32
  • $\begingroup$ I would say that even if your metrics are not getting worse (such as in your case, right plot), but the loss is getting worse (val loss, left plot), then there is probably overfitting going on. It might not have much of an effect on your val data at this moment, but the effect might show when you test it on another set. $\endgroup$
    – user2974951
    Commented Jun 10, 2021 at 12:48

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.