.PB

Introduction

SavedModel is the standard way to save a complete model in TensorFlow. It is a folder, not a single file: a saved_model.pb file describing the model’s computation, beside folders for its weights and any extra files it needs. Because it holds the graph as well as the weights, it can be loaded and run without the Python code that built the model, which is how TensorFlow Serving, the TensorFlow Lite converter and TensorFlow.js take models in.

Structure

saved_model.pb is a binary Protocol Buffers message of type SavedModel, holding one or more MetaGraphDefs. Each has a graph of operations, tags such as serve, and named signatures, usually serving_default, that give the inputs and outputs a caller should use. A text version of the same message may be saved as saved_model.pbtxt. The weights are stored in the same layout as a TensorFlow checkpoint.

Older TensorFlow 1 code also wrote a frozen graph: a single .pb file holding a GraphDef with the weights folded in as constants.

Also used for

• Protocol Buffers binary data .pb

Links