Adding different loss to tf.estimator.Head

[le-dawg]

Hi all,

I seek to optimize an ensemble for binary classification. My established baseline uses the binary_crossentropy loss provided by keras. Using the same notation yields unsupported callable, because it seems the base_head.Head()does not tie into the default tf implementation.

What can I do train with the binary crossetropy?

[cweill]

Have you tried tf.estimator.BinaryClassHead()? I believe it uses the same loss under the hood, specifically sigmoid_cross_entropy.

[le-dawg]

Well, I figured that out in the meantime.
But a problem persists:
BinaryClassHead It uses the correct loss but when I predict using a simple_dnn AdaNet of three iterations via estimator.predict() the network always predicts with class 0.

The same head works on the canned tf.estimator.LinearClassifier which I would suspect to be a problem regarding an incorrect loss function. I can't troubleshoot the AdaNet estimator any deeper than this. I will take any help I can!