What does this results say?

Part 1 (2017)
1

In simple way, In 100 epochs for mnist, for 37800 training values and 4200 Validation values,

train_loss : 0.0112 train_acc : 0.9975
val_loss : 0.0478 val_acc : 0.9927

These are the best values that are saved using ModelCheckpoint in 77/100 Epoch. This is the best one saved.

1 ) Is it overfitting ?
2)If yes by how much ?
3) Is this much overfitting ok ?

2

  1. By definition it is overfitting if the validation loss is bigger than the training loss.

  2. Not much if you have 99.27% accuracy

  3. Depends on the application. 99.27% is pretty good for recognising a digit. Less good if it is the chance of your plane landing safely.

1 Like
3

@simoneva Trying to confirm,

  1. Is it ok to go ahead and save the model weights for these results for practical application of it ?
  2. Is there any measure or standard rules for how much epoch to run for a model ?
  3. Is there any threshold for number of epochs to run that would suffice for any model?

Thank you for your previous answer. It makes sense now.