Keras Autotuner Demo

How did the autotuner return 48 as the best value for the hidden dense layer?
Earlier in the video, the accuracy was 0.9866 using a dense layer of 512 units and with 48 units it only has an accuracy of 0.9564. Shouldn’t the tuner maximise the accuracy? or is it maximising other metrics as well?

Hi @remimse
Please consider that the autotuner has been configured with an early stopping callback with a ‘patience’ = 5. Robert says that
“which means that it (the validation loss) doesn’t change significantly, or if it doesn’t change significantly in five epics, then stop searching on this iteration”
So the autotuner stopped when this condition has been met. Anyway when the model has been retrained with 48 units the accuracy is really high (a bit less than the previous one) even if with only 48 units. Robert remarks that the train process is now 3 times quicker than the previous one and this improvement has been achieved using 48 units instead of 512 of course.

hope this can help