how to load tf2 keras model ?

[GerScau]

https://github.com/tensorflow/java/issues/109 it seems tensorflow java give up to merge keras module.

here is a java developer, i used to using session/graph to load tf1 models generate by python. But when i turnto tf2, it drop the session/graph module, so i turn to use keras to trains models.

now i cannt find any java method on tensorflow/java to load the keras models, also i wonther why session/graph class still could be found on this project, which is drop by tf2 ??

i think there would be people like me, who just want to use java api to use tf2 models, but trains model by python.
appreciate for any help

[zaleslaw]

@GerScau What format for Keras model saving are you going to use? h5? json config or yaml?
One way now is loading graph and variables for inference purposes.
@karllessard correct me if I'm wrong

[aday00]

If your Keras model is saved in the latest protobuf format (so you have a *.pb file for your model, and a directory of other metadata from Kera's saved model), which I think is default for TF2, you want to use SavedModelBundle in TensorflowJava to load the Keras model.

I use TensorflowJava from Scala like this, but Java is similar:

val saved_model: org.tensorflow.SavedModelBundle = org.tensorflow.SavedModelBundle.load(your_pretrained_model_dirname, "serve")

For Java code, read the excellent SavedModelBundleTest.java, which uses SavedModelBundle a lot. https://github.com/tensorflow/java/blob/master/tensorflow-core/tensorflow-core-api/src/test/java/org/tensorflow/SavedModelBundleTest.java

[karllessard]

@zaleslaw , I don't think there is a limitation for loading a model for training but I've also only did it for inference so far.

@GerScau The SavedModelBundle only support TF saved model for now, not h5. One known limitation though is that TF Java only support "session centric" models (i.e. these models that only have a single graph), like estimators do, please see more details at the bottom of this issue.