TensorSpace.js

Present Tensor in Space

English | 中文

TensorSpace is a neural network 3D visualization framework built using TensorFlow.js, Three.js and Tween.js. TensorSpace provides Keras-like APIs to build deep learning layers, load pre-trained models, and generate a 3D visualization in the browser. From TensorSpace, it is intuitive to learn what the model structure is, how the model is trained and how the model predicts the results based on the intermediate information. After preprocessing the model, TensorSpace supports to visualize pre-trained model from TensorFlow, Keras and TensorFlow.js.

Fig. 1 - Interactive LeNet created by TensorSpace

Table of Content

Motivation

TensorSpace is a neural network 3D visualization framework designed for not only showing the basic model structure, but also presenting the processes of internal feature abstractions, intermediate data manipulations and final inference generations.

By applying TensorSpace API, it is more intuitive to visualize and understand any pre-trained models built by TensorFlow, Keras, TensorFlow.js, etc. TensorSpace introduces a way for front end developers to be involved in the deep learning ecosystem. As an open source library, TensorSpace team welcomes any further development on visualization applications.

Interactive -- Use Layer API to build interactive model in browsers.
Intuitive -- Visualize the information from intermediate inferences.
Integrative -- Support pre-trained models from TensorFlow, Keras, TensorFlow.js.

Getting Started

Fig. 2 - TensorSpace Workflow

1. Install TensorSpace

Install in the Basic Case

Step 1: Download Dependencies

Download dependencies build files TensorFlow.js (tf.min.js), Three.js (three.min.js), Tween.js (tween.min.js), TrackballControls (TrackballControls.js).

Step 2: Download TensorSpace

Download TensorSpace build file tensorspace.min.js from Github, NPM, TensorSpace official website or CDN:

<!-- Replace "VERSION" with the version you want to use. -->
<script src="https://cdn.jsdelivr.net/npm/tensorspace@VERSION/dist/tensorspace.min.js"></script>

Step 3: Include Build Files

Include all build files in web page.

<script src="tf.min.js"></script>
<script src="three.min.js"></script>
<script src="tween.min.js"></script>
<script src="TrackballControls.js"></script>
<script src="tensorspace.min.js"></script>

Install in the Progressive Framework

Step 1: Install TensorSpace
- Option 1: NPM
```
npm install tensorspace
```
- Option 2: Yarn
```
yarn add tensorspace
```
Step 2: Use TensorSpace

import * as TSP from 'tensorspace';

Checkout this Angular example for more information.

2. Preprocess the Pre-trained Model

Before applying TensorSpace to visualize the pre-trained model, there is an important pipeline - TensorSpace model preprocessing ( Checkout this article for more information about TensorSpace preprocessing ). We can use TensorSpace Converter to quickly complete the TensorSpace Preprocessing.

For example, if we have a tf.keras model in hand, we can use the following TensorSpace-Converter conversion script to convert a tf.keras model to the TensorSpace compatible format:

$ tensorspacejs_converter \
    --input_model_from="tensorflow" \
    --input_model_format="tf_keras" \
    --output_layer_names="padding_1,conv_1,maxpool_1,conv_2,maxpool_2,dense_1,dense_2,softmax" \
    ./PATH/TO/MODEL/tf_keras_model.h5 \
    ./PATH/TO/SAVE/DIR

Note:

Make sure to install tensorspacejs pip package, and setup a TensorSpace-Converter runtime environment before using TensorSpace-Converter to preprocess the pre-trained model.
Based on different training libraries, we provide different preprocessing tutorials: TensorFlow Tutorial, Keras Tutorial, TensorFlow.js Tutorial.
Checkout TensorSpace-Converter Repo for more information about TensorSpace-Converter.

Fig. 3 - TensorSpace-Converter Usage

3. Using TensorSpace to Visualize the Model

If TensorSpace is installed successfully and the pre-trained deep learning model is preprocessed, let's create an interactive 3D TensorSpace model.

For convenience, we will use the the resources from this repository's HelloWorld directory, which includes preprocessed TensorSpace compatible LeNet model and sample input data ("5") as an example to illustrate this step. All source code can be found in helloworld.html.

First, we need to new a TensorSpace model instance:

let container = document.getElementById( "container" );
let model = new TSP.models.Sequential( container );

Next, based on the LeNet structure: Input + Padding2D + 2 X (Conv2D & Maxpooling) + 3 X (Dense), build the Topology of the TensorSpace model:

model.add( new TSP.layers.GreyscaleInput() );
model.add( new TSP.layers.Padding2d() );
model.add( new TSP.layers.Conv2d() );
model.add( new TSP.layers.Pooling2d() );
model.add( new TSP.layers.Conv2d() );
model.add( new TSP.layers.Pooling2d() );
model.add( new TSP.layers.Dense() );
model.add( new TSP.layers.Dense() );
model.add( new TSP.layers.Output1d({
    outputs: ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
}) );

Last, we should load our preprocessed TensorSpace compatible model and use init() method to create the TensorSpace model:

model.load({
    type: "tensorflow",
    url: './PATH/TO/MODEL/model.json'
});
model.init(function(){
    console.log("Hello World from TensorSpace!");
});

We can get the following Fig. 3 model in the browser if everything looks good.

Fig. 4 - LeNet model without any input data

We provide a extracted file which is a handwritten "5" as the input of our model: (online demo)

model.init(function() {
    model.predict( image_5 );
});

We put the predict( image_5 ) method in the callback function of init() to ensure the prediction is after the initialization complete.

Click the CodePen logo to try it in CodePen:

Fig. 5 - LeNet model with input data "5"

Example

LeNet [ TensorFlow.js model ]

➡ Live Demo

Fig. 6 - Interactive LeNet created by TensorSpace

AlexNet [ TensorFlow model ]

➡ Live Demo

Fig. 7 - Interactive AlexNet created by TensorSpace

Yolov2-tiny [ TensorFlow model ]

➡ Live Demo

Fig. 8 - Interactive Yolov2-tiny created by TensorSpace

ResNet-50 [ Keras model ]

➡ Live Demo

Fig. 9 - Interactive ResNet-50 created by TensorSpace

Vgg16 [ Keras model ]

➡ Live Demo

Fig. 10 - Interactive Vgg16 created by TensorSpace

ACGAN [ Keras model ]

➡ Live Demo

Fig. 11 - Interactive ACGAN created by TensorSpace

MobileNetv1 [ Keras model ]

➡ Live Demo

Fig. 12 - Interactive MobileNetv1 created by TensorSpace

Inceptionv3 [ Keras model ]

➡ Live Demo

Fig. 13 - Interactive Inceptionv3 created by TensorSpace

LeNet Training Visualization [ TensorFlow.js dynamic model ]

Visualize the LeNet Training Process with TensorSpace.js and TensorFlow.js.

➡ Live Demo

Fig. 14 - LeNet Training 3D Visualization

View models locally

As some models above are extremely large, view them locally may be a good choice.

Step 1: clone TensorSpace Repo

git clone https://github.com/tensorspace-team/tensorspace.git

Step 2:

Open "html" file in examples folder in local web server.

Documentation

For a quick start, checkout Getting Start
To learn more about the Basic Concepts
To process a deep learning model, checkout Model Preprocessing, TensorSpace-Converter
To learn core components: Models, Layers and Merge Function
Checkout the official website TensorSpace.org for more about TensorSpace.