See the following link

https://github.com/keras-team/keras/issues/2548

Best Regards Hunar A. Ahmed

Dear Ahmed,

It would be more meaningful to compute the average accuracy across the epochs where the accuracies consistently high during the model's training rather than choosing the best accuracy at certain epoch. Such an average accuracy would better reflect the actual performance of your trained model when tested. In other words, if you observed that the accuracy is steadily high over several epochs, you could stop the training within such epochs and compute the average accuracy to avoid over-fitting.

I hope you would find this tips helpful.

Good luck.

Dear Hunar A. Ahmed ,

You can select the epoch where the model achieved the best accuracy. In addition, you can stop the training when the accuracy is not improving, given a threshold.

See ModelCheckpoint and EarlyStopping topics in https://keras.io/callbacks/

I hope this helps you.

which one is better in my case, early stopping or average? I think early stopping is not real accuracy because in some runs, in one or two epochs I get 100% accuracy?

I think you need to select a model (i.e., your model in epoch n) and report its final accuracy. I'm not sure if taking the mean accuracy over epochs makes sense.

Accuracy during training is not a good measure of your final model but is for validation purpose and to show if you are overfitting or underfitting. You must test your model at the end using new unseen data to get your accuracy.

e.g in keras;

model.evaluate(test_data)

hello

in this forum you can find hints related to this question:

How can I get both test accuracy and validation accuracy for each epoch

https://machinelearningmastery.com/display-deep-learning-model-training-history-in-keras/