Make it possible to use TensorFlow models as elements of a FlinkML pipeline (as described here). Possible features include:

• Use TF models as FlinkML transformers and predictors
• Support FlinkML datatypes
• Train new models to be used for inference and/or exported

• Integrate graph state into Flink operator state
• Support checkpointing
• Support keyed streams

Can you please help me to get the maven dependency for flink-tensorflow_2.11? I could not find the dependency with this name in any repository.

TensorFlow now publishes a library containing the generated proto classes. The FTF library should take a dependency on that in lieu of the generated classes it now contains.

https://search.maven.org/#search%7Cga%7C1%7Ca%3A%22proto%22

http://javadoc.io/doc/org.tensorflow/proto/

Hi, I am a new Flink learner. These days I encounter a problem in your project. after successfully testing the building by "mvn surefire:test", I try to run the Johnny model example. When I try to run the example in cluster by flink run -c org.apache.flink.contrib.tensorflow.examples.inception.Johnny flink-tensorflow-examples/target/flink-tensorflow-examples_2.11-0.1-SNAPSHOT.jar /path/to/inception5h /path/to/photos, the execution job failed with the following error:

java.lang.NoSuchMethodError: scala.Predef$.ArrowAssoc(Ljava/lang/Object;)Ljava/lang/Object;

If you have spare time, It would be very nice of you to give me some tips.
Thanks.

Let's prove that serving TF models in-process has a performance advantage over making RPC calls to TF Serving.

I can't find any flink-tensorflow-related jar in maven repo。
I'd like to use tensorflow-trained model in flink. Could help me to install this?

ps: The doc should be enhanced

Make it possible to define TF graphs from scratch, in addition to existing support for using externally-developed models.

Because the object model of TensorFlow graphs is extensive and ever-growing, we should hesitate to introduce a Flink-specific DSL. We should consider creating or incorporating a pure Scala TensorFlow library where such a DSL could be developed.

The library today has numerous layers of functionality that could benefit from refactoring, to reduce dependencies and clarify concepts. The conceptual layering (from bottom up):

1. Scala language bindings
   a. focus on using the TensorFlow library in Scala (not Flink specific)
2. Core Flink integration code
   a. graph/model loaders
   b. tensors in the Flink type system (e.g. TensorValue, type tags)
   c. core lifecycle classes (e.g. ModelAwareFunction)
   d. (future) integration of graph state with Flink managed state
3. ML specializations
   a. representations of specific ML signatures ('classify', 'regress', 'predict')
   b. TF Serving interoperability
   c. FlinkML integration
   d. converters for specific datatypes (e.g. LabeledVector, Example)

Ideally the core module (2) would shed its dependency on 'twitter-bijection', though it may still be useful in (3).

I compiled FTF lib and examples，and imported lib into examples project.However , eclipse showed “The import org.tensorflow.framework cannot be resolved.object framework is not a member of package org.tensorflow.” exp:

import org.tensorflow.framework.{SignatureDef, TensorInfo}

and I also could not find org.tensorflow.framework

                                                                                 Thanks

Implement out-of-box interoperability with TF Serving models. Given a model that works in TF Serving, make it easy to use directly in a Flink program, e.g.:

• implementations of "predict" and "classification" signature
• MapFunction implementations w/ appropriate types (tf.Example, Array[Byte])
• Example code

Some snippets to look at:

• Serving a TensorFlow Model
• mnist_saved_model.py
• API definitions