Still overfitting on Horse or Humans dataset


I try without success to overcome overfitting with the horse or humans dataset.
I’m not able to have a proper graph for accuracy on validation dataset.
Basically, I have something like this

I tried:

  • Data Augmentation
  • Dropout
  • Smaller or larger Batch size
  • Smaller network
  • Optimizer change
  • Learning rate change

it’s still oscillating and I don’t know what to do. I would like to have some insights for having something smoother an a validation accuracy that follow the training accuracy.

This is my model so far:

[code removed - moderator]

Any helps ? Thank

Given that the training accuracy is far better than validation accuracy, please try more augmentations.
See this topic as well.

Thank ! When you say more augmentations, what does that mean @balaji.ambresh ? rise the parameters ?

Please see the augmentations that are available in ImageDataGenerator. Shear, shift & zoom are some transformations that you can consider. The range of values also play an important role on how much room you provide to augment data.