how to do early stopping while training adanet

[guoqingyu319]

When using adanet on my own data,I face a overfitting problem. I have to try to add early stopping strategy while training adanet.My problem is early stopping strategy stop the adanet iteration.So How can I do early stopping in each iteration so that I can solve the overfitting problem.
1.I set 8 iterations,each iteration has 200000 step
2.The whole training stop at 160000 step in first iteration,

[cweill]

@guoqingyu319: We don't yet support early stopping per iteration, though it's something we would like to add. For now, the best thing is to set the max_iteration_steps to the right number to prevent overfitting (though sometime a little overfitting is good for the ensemble performance).

[guoqingyu319]

@cweill thanks a lot.In our scenario,we wanna use adanet to do automl.The biggest problem is some hyperparameters in adanet such as max_iteration_steps can not be automated completly.Different datasets have different max_iteration_steps we need.So the right number you said is also a parameter we should try.Thank you so much!

[cweill]

@guoqingyu319 What we generally do is tune the train steps (i.e. number of epochs) to find the best AdaNet model. So your solution sounds like the way to go. :)