Frameworks like tensorflow should also be used in production. Tensorflow is written in highly optimized C++ and/or Cuda and will thus easily outperform anything you will write yourself. There is also a large amount of inbuilt tools for monitoring, load balancing etc. that comes with tensorflow.
If you run your system on your machine you can use any language as a wrapper. If you really think you need a level of optimization for small tasks such as preprocessing etc. that can not be done in Python, just use the C++ API for tensorflow. Just let tensorflow (or torch or whatever production lvl framework you are using) do the heavy lifting of handling the neural network.
For 99.99% percent of tasks Python(using numpy, tf, dask etc.) will be just fine.
For large scale applications you can host your model on aws/azure/google cloud etc. They offer optimized services for hosting tensorflow models (eg. as REST endpoints). Lambda fucntions/Azure Functions can be a great tool to do the preprocessing step in a costefficient and scaleable way.
与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…