Could provide interface of tensorflow old api rather than use estimator ?

[svjack]

I can understand the motivation of use estimator api to
standardize the model construct and train methods in configuration,
But It may be complicated to debug with estimator api.

[cweill]

Can you please give an example of what you would prefer the API to look like? Perhaps the new adanet.AutoEnsembleEstimator may be a better API for your needs? I'm considering a simpler Keras-base API as well. But other options are welcome.

[bhack]

https://github.com/tensorflow/estimator/issues/4

[cweill]

My understanding is tf.Estimator will continue to be the recommended way to train and serve models in TF 2.0.

[svjack]

My understanding is tf.Estimator will continue to be the recommended way to train and serve models in TF 2.0.

yes, for the compatibility conveyed by canned