I would like to run tensorflow serving without the HTTP or gRPC interfaces. In fact, I don't want it to even bind a port for receiving client requests. I'd like to leverage it as a library directly within my existing process, which is a python microservice. I know that it has a C++ API - so it seems possible that those APIs could be exposed to Python and it could be leveraged directly. Anyone tried that approach?
1 Answers
We ended up simply running tensorflow-serving as a subprocess in a microservice app based on python/flask. We wrote up a blog post describing the implementation: https://medium.com/p/78d422d10c2c
Related