@tensorflow/tfjs-converter

What is @tensorflow/tfjs-converter?

@tensorflow/tfjs-converter is a library that allows you to import TensorFlow SavedModel and Keras models into TensorFlow.js. This enables you to run pre-trained models in the browser or in Node.js, making it easier to deploy machine learning models in web applications.

What are @tensorflow/tfjs-converter's main functionalities?

Import TensorFlow SavedModel

This feature allows you to import a TensorFlow SavedModel into TensorFlow.js. The code sample demonstrates how to load a SavedModel from a specified path and log the model to the console.

const tf = require('@tensorflow/tfjs');
const tfConverter = require('@tensorflow/tfjs-converter');

async function loadModel() {
  const model = await tfConverter.loadGraphModel('path/to/saved_model');
  console.log(model);
}

loadModel();

Import Keras Model

This feature allows you to import a Keras model into TensorFlow.js. The code sample demonstrates how to load a Keras model from a specified JSON file and log the model to the console.

const tf = require('@tensorflow/tfjs');
const tfConverter = require('@tensorflow/tfjs-converter');

async function loadKerasModel() {
  const model = await tfConverter.loadLayersModel('path/to/keras_model.json');
  console.log(model);
}

loadKerasModel();

Run Inference

This feature allows you to run inference using a loaded model. The code sample demonstrates how to load a SavedModel, create a tensor as input, run the model's predict function, and print the output.

const tf = require('@tensorflow/tfjs');
const tfConverter = require('@tensorflow/tfjs-converter');

async function runInference() {
  const model = await tfConverter.loadGraphModel('path/to/saved_model');
  const input = tf.tensor([1, 2, 3, 4]);
  const output = model.predict(input);
  output.print();
}

runInference();

Other packages similar to @tensorflow/tfjs-converter

@tensorflow/tfjs

@tensorflow/tfjs is the core TensorFlow.js library that provides the main functionalities for defining, training, and running machine learning models in the browser and Node.js. While @tensorflow/tfjs-converter focuses on importing pre-trained models, @tensorflow/tfjs provides the tools to build and train models from scratch.

onnxjs

onnxjs is a library for running ONNX (Open Neural Network Exchange) models in the browser and Node.js. It provides similar functionalities to @tensorflow/tfjs-converter but focuses on the ONNX model format instead of TensorFlow SavedModel and Keras models.

brain.js

brain.js is a JavaScript library for neural networks that runs in the browser and Node.js. It provides functionalities for creating, training, and running neural networks, but it does not support importing TensorFlow or Keras models like @tensorflow/tfjs-converter.

README

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:

Usage:

SavedModel example:

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

Session bundle model example:

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

Tensorflow Hub module example:

$ tensorflowjs_converter \
    --input_format=tf_hub \
    'https://tfhub.dev/google/imagenet/mobilenet_v1_100_224/classification/1' \
    /mobilenet/web_model

Keras h5 model example:

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

tf.keras SavedModel model example:

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

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, tfjs_layers_model 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 tfjs_layers_model, tfjs_graph_model 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 https://www.tensorflow.org/hub/common_signatures/.
--strip_debug_ops	Strips out TensorFlow debug operations Print, Assert, CheckNumerics. Defaults to True.
--quantization_bytes	How many bytes to optionally quantize/compress the weights to. Valid values are 1 and 2. which will quantize int32 and float32 to 1 or 2 bytes respectively. The default (unquantized) size is 4 bytes.

Format Conversion Support Tables

Note: Unless stated otherwise, we can infer the value of --output_format from the value of --input_format. So the --output_format flag can be omitted in most cases.

Python-to-JavaScript

--input_format	--output_format	Description
keras	tfjs_layers_model	Convert a keras or tf.keras HDF5 model file to TensorFlow.js Layers model format. Use tf.lodLayersModel() to load the model in JavaScript.
keras_saved_model	tfjs_layers_model	Convert a tf.keras SavedModel model file (from tf.contrib.saved_model.save_keras_model) to TensorFlow.js Layers model format. Use tf.lodLayersModel() to load the model in JavaScript.
tf_frozen_model	tfjs_graph_model	Convert a TensorFlow Frozen Graph (.pb) file to TensorFlow.js graph model format. Use tf.loadGraphModel() to load the converted model in JavaScript.
tf_hub	tfjs_graph_model	Convert a TF-Hub model file to TensorFlow.js graph model format. Use tf.loadGraphModel() to load the converted model in JavaScript.
tf_saved_model	tfjs_graph_model	Convert a TensorFlow SavedModel to TensorFlow.js graph model format. Use tf.loadGraphModel() to load the converted model in JavaScript.
tf_session_bundle	tfjs_graph_model	Convert a TensorFlow Session Bundle to TensorFlow.js graph model format. Use tf.loadGraphModel() to load the converted model in JavaScript.

JavaScript-to-Python

--input_format	--output_format	Description
tfjs_layers_model	keras	Convert a TensorFlow.js Layers model (JSON + binary weight file(s)) to a Keras HDF5 model file. Use keras.model.load_model() or tf.keras.models.load_model() to load the converted model in Python.

Web-friendly format

The conversion script above produces 2 types of files:

• model.json (the dataflow graph and weight manifest file)
• group1-shard\*of\* (collection of binary weight files)

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

  https://storage.cloud.google.com/tfjs-models/savedmodel/mobilenet_v1_1.0_224/model.json
  https://storage.cloud.google.com/tfjs-models/savedmodel/mobilenet_v1_1.0_224/group1-shard1of5
  ...
  https://storage.cloud.google.com/tfjs-models/savedmodel/mobilenet_v1_1.0_224/group1-shard5of5

Step 2: Loading and running in the browser

Instantiate the GraphModel class and run inference.

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

const MODEL_URL = 'https://.../mobilenet/model.json';

const model = await tf.loadGraphModel(MODEL_URL);
const cat = document.getElementById('cat');
model.execute({input: tf.browser.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 loadGraphModel(MODEL_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 loadGraphModel with the path to the model files:

// Load the tfjs-node binding
import * as tf from '@tensorflow/tfjs-node';

const MODEL_PATH = 'file:///tmp/mobilenet/model.json';
const model = await tf.loadGraphModel(MODEL_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 modelUrl = "https://example.org/model/model.json";

const model = await fetch(modelUrl);
this.weightManifest = (await model.json())['weightsManifest'];
const weightMap = await tf.io.loadWeights(
        this.weightManifest, "https://example.org/model");

FAQ

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 also 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. Can I quantize the weights over the wire?

Yes, you can use the --quantization_bytes option to compress int32/float32 to 1 or 2 bytes. Here is an example of 8-bit quantization:

tensorflowjs_converter \
    --input_format=tf_hub \
    --quantization_bytes=1
    'https://tfhub.dev/google/imagenet/mobilenet_v1_100_224/classification/1' \
    /mobilenet/web_model

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.

Development

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

$ git clone https://github.com/tensorflow/tfjs-converter.git
$ 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