The final step of the machine learning workflow is the deployment to production. In this phase, we want the trained model to be deployed on a device, but more than often the device has an entirely different runtime environment with respect to the one used during training.

TensorFlow, thanks to its SavedModel serialization format, allows deploying a trained model to several “deployment platforms”. Your model can run on a browser, in a Java application, in a Python script, and last but not least on every device that can run a C program. There is, in fact, a TensorFlow C API that can also be used for generating language bindings – and that is where Go, with its FFI for the C language, jumps in. In this talk, we will learn the basics of the TensorFlow Go bindings, their limitations, and how the tfgo library simplifies their usage.
Moreover, the flexibility of the SavedModel serialization format is presented, and we will be able to design a deployment environment for incremental learning – in Go!