1
$\begingroup$

I'm currently training a CNN to do a binary classification. I'm getting fairly good results, but unfortunately the training is very unstable. Just by changing the seed the relative error changes by 20-30%.

What can be the cause of this and how can I prevent it?

Other info:

  • The amount of data is rather small.
  • I'm starting from imagenet snapshot.
$\endgroup$
  • $\begingroup$ Are you using tensorflow? You may make a new session for each for. It was one of my main issues :) $\endgroup$ – Media Feb 21 at 20:51
1
$\begingroup$
  • Understand a data
  • Find the most suitable metrics
  • handle missing values
  • normalize, scale etc your variables
  • feature selecting
  • use cross validation(kfold for example)
  • try oversampling or undersampling
  • trying other models
| improve this answer | |
$\endgroup$
1
$\begingroup$

I would recommend to read the following:

http://karpathy.github.io/2019/04/25/recipe/

| improve this answer | |
$\endgroup$

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.