.PB File Format – The SavedModel folder

TensorFlow SavedModel (saved_model.pb)

A SavedModel is a folder, not a single file. saved_model.pb describes the computation; the trained weights and any extra files sit beside it. Only saved_model.pb is this format; the other files have formats of their own.

What the folder holds

  1. /saved_model.pb the SavedModel message: meta graphs, signatures and functionsfile
  2. /variables/ variables.index and variables.data-00000-of-00001 (one or more data shards): the values of the variables, in the layout of a TensorFlow checkpointfolder
  3. /assets/ extra files the graph reads, such as vocabularies; each is named by an AssetFileDeffolder
  4. /fingerprint.pb hashes that identify the model, written by newer versions of TensorFlowfile
The contents of a SavedModel folder

The SaverDef of each meta graph names the ops that restore the variables from variables/.

saved_model.pbtxt

The same SavedModel message may be saved in Protocol Buffers text format as saved_model.pbtxt instead. It is printable text, so the explorer does not read it as this format.

A Git LFS pointer in place of the file

Models kept in Git repositories often store large files with Git LFS. When such a repository is cloned without fetching the LFS files, saved_model.pb is not the model but a short text file that begins:

version https://git-lfs.github.com/spec/v1
oid sha256:…
size …
A Git LFS pointer named saved_model.pb

The explorer turns such a file down, because it is printable text. Fetch the LFS files (git lfs pull) to get the real model.

History

TensorFlow was released as open source in November 2015, with GraphDef already in place. The release notes of TensorFlow 1.0.0 say “saved_model is available as tf.saved_model”; 1.1.0 let Java load SavedModels, and 1.2.0 added the saved_model_cli tool. TensorFlow 2, released in 2019, made SavedModel the standard way to save a whole model. saved_model_schema_version has been 1 throughout.