Jan Vladimir Mostert

Idea Incubator

Nov 13, 2015

You might have some luck using Jython, Tensorflow is in Python, so if you want to run it on the JVM, Jython might be your best bet.

Alternatively, write a Python HTTP / AMQP API that you can call from Java and write wrappers for each function you need ... or write the TensorFlow logic in Python and call the functions from Java using the AMQP / HTTP wrappers.

Jan Vladimir Mostert

Idea Incubator

Nov 18, 2015

What I've done in the past to get NumPy or other libraries working without having to install them is to include the source of those libraries in my project. With NumPy it might be tricky due to its C-bindings.

JNI + C++ sounds like the easier option here as Killian suggested or writing AMQP bindings and calling those methods via RabbitMQ.

Kolja Bailly

Nov 17, 2015

As far as i tested, using jython with tensorflow fails because you cannot install the required module numPy: scipy.org/scipylib/faq.html

Jan Vladimir Mostert

Idea Incubator

Nov 14, 2015

I was more thinking in terms of using Jython to import those python bindings to Java, didn't know there was a full blown C++ API. If portability is not an issue, then JNI is definitely an option.

killian

attracted by deep learning, natural language processing and dev web

Nov 14, 2015

Actually TensorFlow is mostly made in C++ with python bindings , I think don't Jython is able to import c++ based library. Using JNI with the C++ API can be a possibility to. Apparently Google planned to add various language support using SWIG tensorflow.org/api_docs

Jan Vladimir Mostert

Idea Incubator

Nov 14, 2015

I usually do RPC over AMQP, it's very fast. gRPC looks like an interesting alternative.

Shrivatsan Rajagopalan

learner @ <3

Nov 14, 2015

Thrift Protobuff gRPC could be better options as HTTP would be moving bytes up and down the OS stacks.

AMQP is binary over wire => +1 HTTP/1.1 => so-so

Search Hashnode