Convert TensorFlow SavedModel and Keras models to TensorFlow.js
pyu10055 increase the converter version to 0.6.7 (#251)
* increase the version to 0.6.7

* add the missing python version file
Latest commit 4b79271 Nov 15, 2018

Build Status

Getting started

TensorFlow.js converter is an open source library to load a pretrained TensorFlow SavedModel, Frozen Model, Session Bundle or TensorFlow Hub module into the browser and run inference through TensorFlow.js.

(Note: TensorFlow has deprecated session bundle format, please switch to SavedModel.)

A 2-step process to import your model:

  1. A python pip package to convert a TensorFlow SavedModel/Frozen Model/Session Bundle to a web friendly format. If you already have a converted model, or are using an already hosted model (e.g. MobileNet), skip this step.
  2. Javascript API, for loading and running inference.

Step 1: Converting a SavedModel, Keras h5, tf.keras SavedModel, Session Bundle, Frozen Model or TensorFlow Hub module to a web-friendly format

  1. Install the TensorFlow.js pip package:
  $ pip install tensorflowjs
  1. Run the converter script provided by the pip package:


SavedModel example:

$ tensorflowjs_converter \
    --input_format=tf_saved_model \
    --output_node_names='MobilenetV1/Predictions/Reshape_1' \
    --saved_model_tags=serve \
    /mobilenet/saved_model \

Frozen model example:

$ tensorflowjs_converter \
    --input_format=tf_frozen_model \
    --output_node_names='MobilenetV1/Predictions/Reshape_1' \
    --saved_model_tags=serve \
    /mobilenet/frozen_model.pb \

Session bundle model example:

$ tensorflowjs_converter \
    --input_format=tf_session_bundle \
    --output_node_names='MobilenetV1/Predictions/Reshape_1' \
    /mobilenet/session_bundle \

Tensorflow Hub module example:

$ tensorflowjs_converter \
    --input_format=tf_hub \
    '' \

Keras h5 model example:

$ tensorflowjs_converter \
    --input_format=keras \
    /tmp/my_keras_model.h5 \

tf.keras SavedModel model example:

$ tensorflowjs_converter \
    --input_format=keras_saved_model \
    /tmp/my_tf_keras_saved_model/1542211770 \

Note that the input path used above is a subfolder that has a Unix epoch time (1542211770) and is generated automatically by tensorflow when it saved a tf.keras model in the SavedModel format.

Positional Arguments Description
input_path Full path of the saved model directory, session bundle directory, frozen model file or TensorFlow Hub module handle or path.
output_path Path for all output artifacts.
Options Description
--input_format The format of input model, use tf_saved_model for SavedModel, tf_frozen_model for frozen model, tf_session_bundle for session bundle, tf_hub for TensorFlow Hub module, tensorflowjs for TensorFlow.js JSON format, and keras for Keras HDF5.
--output_node_names The names of the output nodes, separated by commas.
--output_format The desired output format. Must be tensorflowjs (the default) or keras. Not all pairs of input-output formats are supported. Please file a github issue if your desired input-output pair is not supported.
--saved_model_tags Only applicable to SavedModel conversion. Tags of the MetaGraphDef to load, in comma separated format. Defaults to serve.
--signature_name Only applicable to TensorFlow Hub module conversion, signature to load. Defaults to default. See
--strip_debug_ops Strips out TensorFlow debug operations Print, Assert, CheckNumerics. Defaults to True.

Format conversions support table

input format output tensorflowjs output keras
keras ✔️
tensorflowjs ✔️
tf_frozen_model ✔️
tf_hub ✔️
tf_saved_model ✔️
tf_session_bundle ✔️

Web-friendly format

The conversion script above produces 4 types of files:

  • tensorflowjs_model.pb (the dataflow graph)
  • weights_manifest.json (weight manifest file)
  • model.json (the two above, in a single file)
  • group1-shard\*of\* (collection of binary weight files)

For keras input files, the converter generates model.json and group1-shard\*of\*. For other input formats, it generates the tensorflowjs_model.pb, weights_manifest.json, and group1-shard\*of\*.

For example, here is the MobileNet model converted and served in following location:

Step 2: Loading and running in the browser

Instantiate the FrozenModel class and run inference.

import * as tf from '@tensorflow/tfjs';

const MODEL_URL = 'https://.../mobilenet/tensorflowjs_model.pb';
const WEIGHTS_URL = 'https://.../mobilenet/weights_manifest.json';

const model = await tf.loadFrozenModel(MODEL_URL, WEIGHTS_URL);
const cat = document.getElementById('cat');
model.execute({input: tf.fromPixels(cat)});

Check out our working MobileNet demo.

If your server requests credentials for accessing the model files, you can provide the optional RequestOption param.

const model = await loadFrozenModel(MODEL_URL, WEIGHTS_URL,
    {credentials: 'include'});

Please see fetch() documentation for details.

Native File System

TensorFlow.js can be used from Node.js. See the tfjs-node project for more details. Unlike web browsers, Node.js can access the local file system directly. Therefore, you can load the same frozen model from local file system into a Node.js program running TensorFlow.js. This is done by calling loadFrozenModel with the path to the model files:

// Load the tfjs-node binding
import '@tensorflow/tfjs-node';

const MODEL_PATH = 'file:///tmp/mobilenet/tensorflowjs_model.pb';
const WEIGHTS_PATH = 'file:///tmp/mobilenet/weights_manifest.json';
const model = await tf.loadFrozenModel(MODEL_PATH, WEIGHTS_PATH);

You can also load the remote model files the same way as in browser, but you might need to polyfill the fetch() method.

Supported operations

Currently TensorFlow.js only supports a limited set of TensorFlow Ops. See the full list. If your model uses an unsupported ops, the tensorflowjs_converter script will fail and produce a list of the unsupported ops in your model. Please file issues to let us know what ops you need support with.

Loading the weights only

If you prefer to load the weights only, you can use follow code snippet.

import * as tf from '@tensorflow/tfjs';

const weightManifestUrl = "";

const manifest = await fetch(weightManifestUrl);
this.weightManifest = await manifest.json();
const weightMap = await
        this.weightManifest, "");


  1. What TensorFlow models does the converter currently support?

Image-based models (MobileNet, SqueezeNet, add more if you tested) are the most supported. Models with control flow ops (e.g. RNNs) are not yet supported. The tensorflowjs_converter script will validate the model you have and show a list of unsupported ops in your model. See this list for which ops are currently supported.

  1. Will model with large weights work?

While the browser supports loading 100-500MB models, the page load time, the inference time and the user experience would not be great. We recommend using models that are designed for edge devices (e.g. phones). These models are usually smaller than 30MB.

  1. Will the model and weight files be cached in the browser?

Yes, we are splitting the weights into files of 4MB chunks, which enable the browser to cache them automatically. If the model architecture is less than 4MB (most models are), it will also be cached.

  1. Will it support model with quantization?

Not yet. We are planning to add quantization support soon.

  1. Why is the predict() method for inference so much slower on the first call than the subsequent calls?

The time of first call also includes the compilation time of WebGL shader programs for the model. After the first call the shader programs are cached, which makes the subsequent calls much faster. You can warm up the cache by calling the predict method with an all zero inputs, right after the completion of the model loading.


To build TensorFlow.js converter from source, we need to clone the project and prepare the dev environment:

$ git clone
$ cd tfjs-converter
$ yarn # Installs dependencies.

We recommend using Visual Studio Code for development. Make sure to install TSLint VSCode extension and the npm clang-format 1.2.2 or later with the Clang-Format VSCode extension for auto-formatting.

Before submitting a pull request, make sure the code passes all the tests and is clean of lint errors:

$ yarn test
$ yarn lint

To run a subset of tests and/or on a specific browser:

$ yarn test --browsers=Chrome --grep='execute'
> ...
> Chrome 64.0.3282 (Linux 0.0.0): Executed 39 of 39 SUCCESS (0.129 secs / 0 secs)

To run the tests once and exit the karma process (helpful on Windows):

$ yarn test --single-run

To generate the static js file for GraphDef proto, run following steps:

  1. Generate static js file with comment first, in order to generate typescript definition.
$ node_modules/protobufjs/bin/pbjs -t static-module -w commonjs -o src/data/compiled_api.js --no-create --no-encode --no-verify --no-convert --no-delimited --no-beautify src/data/api.proto

$ node_modules/protobufjs/bin/pbts -o src/data/compiled_api.d.ts src/data/compiled_api.js
  1. Replace the static js file with the version without comments.
$ node_modules/protobufjs/bin/pbjs -t static-module -w commonjs -o src/data/compiled_api.js --no-create --no-encode --no-verify --no-convert --no-delimited --no-beautify --no-comments src/data/api.proto