Sep 6, 2018 · 1 min read
Actually, I'm already using a prediction (not a classification!) method. So on the server side, you do not need any change. Just take your model, export it into TF estimator and save it for serving. You need to change a client then — because of a different output signature.
If you use pure TensorFlow without Keras, you may have a look at my sample implementation as I described here or in this great post. I hope, it will be more clear how to prepare the model for serving with different types of the output.
