Tensorflow export and reuse Estimator object in Python

Question

The question is about Tensorflow and its higher API Estimator and more generally about exporting and reusing models.

There exists several facilities to export a tensorflow models: tf.train.Saver, write and importing graph directly with their associated weights.

My interests lies in the tf.estimator.Estimator object and how we could reuse them. My use case is pretty simple: I fit a simple model say in tensorflow and would like to use its predictions in a more involved one.

There exists a export_savemodel method for serving the model, however I would need to make communication in my model through ports and I am not sure if this is optimal for training.

My question are the following:

1. Is there a (simple) way to use and import the output of export_savemodel in python? I guessed you could technically use the C++ code and create wrapper around them?

2. Is there an efficient way to compile an tf.estimator.Estimator and reuse it in Python? The XLA compiler seems be interesting, but I could not find out if I could use the result back in Python.

3. Generally, what are the best practice on how to reuse previously trained models in Tensorflow?

Thanks a lot for your reading!

EDIT (my own answer): The accepted answer offers to use the saved_model API. However, I think it is worthwhile to understand the concept of graph and frozen graph. This gist link illustrates the main steps that are probably replicated in the saved_model api. Basically, you have to:

1. rebuild the graph with redefined input ops;
2. freeze this version;
3. provide the new input to the input_map argument when calling tf.import_graph_def to reload the model backs.

Answer 1

1. Yes, see this doc https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md. In particular, the APIs.Loader.Python section.
2. What do you mean by "compile an estimator"? All the data used by estimator is saved into SavedModel. The reset is simply high-level orchestrating logic. Actual operations like matrix multiplications are provided by the C++ library and run on whatever hardware you have - CPU, GPU, or TPU. XLA is a pretty low-level compiler, far from the Estimator API. For more info about it search for a talk about it "XLA: TensorFlow, Compiled! (TensorFlow Dev Summit 2017)"
3. The link above provides a very high-level API. For the lower layer, see https://www.tensorflow.org/programmers_guide/meta_graph. At even lower layer, there is GraphDef (see links in the meta_graph page)