I'm training two ResNet models on an image dataset. The first one has been trained with random weights, while the other has been pre-trained on ImageNet before.

The second model starts overfitting after 12 epochs for which the training loss is 1.47e-3. On the other hand, the first model does not show clear overfitting behaviour after 70 epochs for which the training loss is 1.17e-3 (hence lower).

Is there a paradox ? How could I explain this ?


To be expected.

Why? Imagenet one will over-fitt faster because it was trained on more than 20k classes, not only that but it learned all of the nuances that it needs to learn, and updating the weights for your new dataset takes smaller amount of time (starts overfitting faster).

On the other hand training NN from scratch takes a lot of time.

| improve this answer | |
  • $\begingroup$ Thanks a lot for the answer $\endgroup$ – Olfa HADDAJI Dec 24 '19 at 14:38

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.