Getting bad results for a classification problem

(tarun) #1

HI folks,

I am trying to create a classifier to classify fighter planes and passenger planes.

I followed the steps mentioned in lesson 2 and found that with default settings the error rate is around 10%

After that I ran the learning rate finder and saw that my graph is pretty different from what I usually see in the lectures. What could be the reason for that. Also I tried different values of the epochs when doing fit_one_cycle, following the longest slope fundamental, from the image it seems 3e-4 - 3e-3 is the best and longest slope. But I am getting worse result as compared to 3e-5, 3e-4 why is that ?

Even after doing fit_one_cycle the best I could get was 8% error rate.

Confusion matrix :

Then I tried doing the image clean up as mentioned in the lesson but I see most of the images are correct.

What could be the reason for the bad performance here ?

Are these classes too close to be compared ?

Please suggest.

PS : Please let me know if there is a better channel to ask such questions.