[Question] How to decide the value of "complexity" param in Subnetwork ?

[svjack]

The paper say: "standard deviations of the outputs of the last hidden layer on the training data as surrogate for Rademacher complexities",
And customizing_adanet.ipynb example use tf.constant(1), is the value choose is robust to the conclusion ?
Or should set the value by standard deviations ?

[cweill]

@svjack: We may open source some of our complexity measures including Standard Deviation. Currently the one we have in adanet_objective.ipynb is sqrt(d) where d is the number of hidden layers. For the customizing_adanet.ipynb, we use a constant complexity since there is only one candidate at each iteration and they all have the same hyperparameters and therefore complexity.