1
$\begingroup$

I have very interesting experience in my CNN binary image classification. Do you think the result is by chance or there is a logic behind it?

I used InceptionV3 transfer with softmax (I know you will say why not ReLU) but it is what I did.

I trained on 100 epochs. But the result was terrible. from the training process I noticed in the 12-th epochs the result is excellent (both train accuracy and validation accuracy ). So I trained the model on 12-th epochs. and suprisely the result on the test data was also excellent.

Does it mean on a few epoch the result is better?

Improve this question
$\endgroup$
1
  • $\begingroup$ What framework are you using? Most frameworks save a model snapshot only when performance is improved. $\endgroup$
    – hH1sG0n3
    Nov 3, 2020 at 14:04

2 Answers 2

Reset to default
2
$\begingroup$

Yes, this is the reason you should use 'early stopping' in your models which will stop training when the model is not improving or you can keep the history of the training to pick the epoch that had the best performance.

The reason you get excellent results in the 12th epoch, but terrible performance in the 100th epoch is simply you are overtraining. By overtraining, you are causing overfitting, and the model is not able to generalize, instead, it imitates your data. Thus, the model will have high accuracy in in-sample data but comparably bad in out-of-sample data when you train a lot.

Moreover, take into account that unnecessarily complex and poorly regularized models are likely to overfit also. Especially, when the input data size is small. But in any case, if you lose performance as you train the model, this is probably because of overtraining.

For this reason, try to always have a graph of training accuracy vs test accuracy (or validation accuracy) by epoch. Thus, you can observe where (in which epochs) train and test accuracy move together. Where your train accuracy is >> test (or validation) accuracy then there happens overfitting, and where your test (or validation) accuracy is >> train accuracy then you are underfitting there.

Improve this answer
$\endgroup$
2
  • $\begingroup$ That is not correct. Model checkpoints are created every time the model achieves a better score. Early stopping functionality has to do with reducing unnecessary runtime from your budget if for example a decline in validation score is noticed for a defined number of epochs. $\endgroup$
    – hH1sG0n3
    Nov 3, 2020 at 14:03
  • $\begingroup$ @hH1sG0n3 thank you for informing. I expressed myself poorly there. Fixed now. $\endgroup$
    – Shahriyar Mammadli
    Nov 3, 2020 at 14:11
0
$\begingroup$

enter image description here

This image perfectly defines your situation, you achieved great results on 12th epoch because after that your model starts to overfit your training data resulting in bad testing results.
12th epoch is your model's Best Fit.

You also would have noticed between 1-12 epochs both your Training as well as Testing error was going down.

Improve this answer
$\endgroup$

Your Answer

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

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