0

I would appreciate your explanation on following :)

I trained CNN network that classifies TWO image classes. I used 'SGD' optimizer and 'categorical_crossentropy' loss function.

My results are as follows: - training loss = 0.28 - training accuracy = 0.93 - validation loss = 0.38 - validation accuracy = 0.82

So it is obvious the model 'overfits' a little (some people say little overfitting is sometimes good).

My question(s) are focused on validation loss and validation accuracy.

Is my network 38% inaccurate or 82% accurate? I`ve read alot about validation loss and validation accuracy, and people say that:

'Unlike accuracy, loss is not a percentage. It is a summation of the errors made for each example in training or validation sets.'

But I still do not intuitively understand the results (38% inaccurate or 82% accurate).

Many thanks! :)

Ivan

Ivan Geek
  • 1
  • 1
  • 1

1 Answers1

0

Your validation set (regarded as the best estimation of real-world performance) is 0.82 meaning it classifies the images correctly with 82% accuracy. The training loss is not a direct function of the accuracy.

csteel
  • 428
  • 1
  • 3
  • 14