Skip to content

Latest commit

 

History

History
53 lines (41 loc) · 1.97 KB

hosting.md

File metadata and controls

53 lines (41 loc) · 1.97 KB
Raw

Hosting your own TensorFlow models

This document describes the HTTP(S)-based protocol implemented by the tensorflow_hub library in order to load TensorFlow models from thub.dev and compatibe services into TensorFlow programs. (This does not cover other types of models that users download directly in the browser, e.g., TFLite.)

Its key feature is to use the same URL in code to load a model and in a browser to view the model documentation.

If you are interested in hosting your own repository of models that are loadable with the tensorflow_hub library, your HTTP(S) distribution service should follow this protocol.

Protocol

When a URL such as https://example.com/model is used to identify a model to load or instantiate, the model resolver will attempt to download a compressed tarball from the URL after appending a query parameter ?tf-hub-format=compressed.

The query param is to be interpreted as a comma separated list of the model formats that the client is interested in. For now only the "compressed" format is defined.

The compressed format indicates that the client expects a tar.gz archive with the model contents. The root of the archive is the root of the model directory and should contain a SavedModel, as in this example:

# Create a compressed model from a SavedModel directory.
$ tar -cz -f model.tar.gz --owner=0 --group=0 -C /tmp/export-model/ .

# Inspect files inside a compressed model
$ tar -tf model.tar.gz
./
./variables/
./variables/variables.data-00000-of-00001
./variables/variables.index
./assets/
./saved_model.pb

Tarballs for use with the deprecated hub.Module() API from TF1 will also contain a ./tfhub_module.pb file. The hub.load() API for TF2 SavedModels ignores such a file.

The tensorflow_hub library expects that model URLs are versioned and that the model content of a given version is immutable, so that it can be cached indefinitely.