Model hosting protocol

This document describes the URL conventions used when hosting all model types on - TFJS, TF Lite and TensorFlow models. It also describes the HTTP(S)-based protocol implemented by the tensorflow_hub library in order to load TensorFlow models from and compatible services into TensorFlow programs.

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

General URL conventions supports the following URL formats:

  • TF Hub publishers follow <a href=""></a><publisher>
  • TF Hub collections follow <a href=""></a><publisher>/collection/<collection_name>
  • TF Hub models have versioned url <a href=""></a><publisher>/<model_name>/<version> and unversioned url <a href=""></a><publisher>/<model_name> that resolves to the latest version of the model.

TF Hub models can be downloaded as compressed assets by appending URL parameters to the model URL. However, the URL parameters required to achieve that depend on the model type:

  • TensorFlow models (both SavedModel and TF1 Hub formats): append ?tf-hub-format=compressed to the TensorFlow model url.
  • TFJS models: append ?tfjs-format=compressed to the TFJS model url to download the compressed or /model.json?tfjs-format=file to read if from remote storage.
  • TF lite models: append ?lite-format=tflite to the TF Lite model url.

For example:

Type Model URL Download type URL param Download URL
TensorFlow (SavedModel, TF1 Hub format) .tar.gz ?tf-hub-format=compressed
TF Lite .tflite ?lite-format=tflite
TF.js .tar.gz ?tfjs-format=compressed

Additionally, some models also are hosted in a format that can be read directly from remote storage without being downloaded. This is especially useful if there is no local storage available, such as running a TF.js model in the browser or loading a SavedModel on Colab. Be conscious that reading models that are hosted remotely without being downloaded locally may increase latency.

Type Model URL Response type URL param Request URL
TensorFlow (SavedModel, TF1 Hub format) String (Path to GCS folder where the uncompressed model is stored) ?tf-hub-format=uncompressed
TF.js .json ?tfjs-format=file

tensorflow_hub library protocol

This section describes how we host models on for use with the tensorflow_hub library. If you want to host your own model repository to work with the tensorflow_hub library, your HTTP(s) distribution service should provide an implementation of this protocol.

Note that this section does not address hosting TF Lite and TFJS models since they are not downloaded via the tensorflow_hub library. For more information on hosting these model types, please check above.

Compressed Hosting

Models are stored on as compressed tar.gz files. By default, the tensorflow_hub library automatically downloads the compressed model. They can also be manually downloaded by appending the ?tf-hub-format=compressed to the model url, for example:


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

Tarballs for use with the legacy TF1 Hub format will also contain a ./tfhub_module.pb file.

When one of tensorflow_hub library model loading APIs is invoked (hub.KerasLayer, hub.load, etc) the library downloads the model, uncompresses the model and caches it locally. 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. Learn more about caching models.

Uncompressed Hosting

When the environment variable TFHUB_MODEL_LOAD_FORMAT or the command-line flag --tfhub_model_load_format is set to UNCOMPRESSED, the model is read directly from remote storage (GCS) instead of being downloaded and uncompressed locally. When this behavior is enabled the library appends ?tf-hub-format=uncompressed to the model URL. That request returns the path to the folder on GCS that contains the uncompressed model files. As an example,
<a href=""></a>
gs://tfhub-modules/google/spice/2/uncompressed in the body of the 303 response. The library then reads the model from that GCS destination.