Hi All,
I was playing around with the lesson 2 code and tried replicating it for cats. Although the model works fine, the error rate is a bit worrisome. It starts off high, decreases after 4 epochs, then increases again. I was under the impression that it should be decreasing after each epoch. Could the image set be an issue or am I doing something wrong?
This is a clear case of overfitting…
It’s important to look closely at the train_loss vs the valid_loss and you can see that the train_loss keeps decreasing while the valid_loss starts to decrease but then increases.
The book/course discusses how to overcome overfitting… Some methods include adding more intense data augmentation or adding regularization (like dropout or weight decay).