Decide on a (or a small number of) model export formats we want to support

[holdenk]

To make these models useful for serving we should have some export support.

[mohamed-ali]

Here are two suggestions:

• PMML : which adds support for predictive models as well as Kmeans, as shown in Spark's documentation
• protobuf: tensorflow supports exporting models into protobuf, moreover, tensor-serving supports deploying these saved models in production. So this format is worth supporting to cover the deep learning models.