1
$\begingroup$

I'm fine tuning with caffenet for a 3 classes classification problem. I have 100 instances of class A, 90 instances of class B and 30 instances of class C. I thought that my net would be biased toward classes A and B but I'm actually getting quite good results. I know that caffenet doesn't take care of imbalance data for me. Maybe it has to do with the fact that my entire train set fits into one batch? Or maybe it has to do with the fact that I'm not really training from scratch but mostly using caffenet's already learned weights? Thanks

$\endgroup$
1
  • $\begingroup$ Your situation is not that imbalanced. When I think of imbalanced I think of ratios closer to 50:1 or 100:1. Think ratio of people with cancer to general population. $\endgroup$
    – kbrose
    May 17, 2018 at 13:43

1 Answer 1

2
$\begingroup$

Are you getting good results on your training or your test set? If it's only the first then you are overfitting, but if it's the second it will not have to do with the fact that your whole training set fits in your batch, the gradients still have a strong pull towards the dominating class(es). But having the pretrained weights will help a lot if the classes are relatively easy to seperate. It will also help if your dataset is similar to the one it was pretrained on, because the intermediate layers that do feature extractions have learned valuable filters for your problem instance.

$\endgroup$
3
  • $\begingroup$ Thank you. I am getting good results on the validation set. I'm doing cross validation BTW. of course I see over fitting on the training at some point because it's such a small data set. $\endgroup$
    – Gil-Mor
    May 28, 2017 at 11:20
  • $\begingroup$ Well then it is the second part, I added a tiny bit more to my answer $\endgroup$ May 28, 2017 at 11:45
  • $\begingroup$ Thanks. I don't know how similar my data set is to ImageNet.. caffenet sees my images as 'coil' or swirl. So maybe it's because the model can't get biased after 80 epochs on 200 images (when it reaches top accuracy)? $\endgroup$
    – Gil-Mor
    May 28, 2017 at 11:54

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.